Remove opts per tick frames

src/dynarmic/src/dynarmic/ir/basic_block.cpp

@@ -39,8 +39,16 @@ Block::iterator Block::PrependNewInst(iterator insertion_point, Opcode opcode, s
     // just pool it!!! - reason why there is an inline buffer is because many small blocks are created
     // with few instructions due to subpar optimisations on other passes... plus branch-heavy code will
     // hugely benefit from the coherency of faster allocations...
-    inlined_inst.emplace_back(opcode);
-    auto* inst = &inlined_inst[inlined_inst.size() - 1];
+    IR::Inst* inst;
+    if (inlined_inst.size() < inlined_inst.max_size()) {
+        inlined_inst.emplace_back(opcode);
+        inst = &inlined_inst[inlined_inst.size() - 1];
+    } else {
+        if (pooled_inst.empty() || pooled_inst.back().size() == pooled_inst.back().max_size())
+            pooled_inst.emplace_back();
+        pooled_inst.back().emplace_back(opcode);
+        inst = &pooled_inst.back()[pooled_inst.back().size() - 1];
+    }
     DEBUG_ASSERT(args.size() == inst->NumArgs());
     std::for_each(args.begin(), args.end(), [&inst, index = size_t(0)](const auto& arg) mutable {
         inst->SetArg(index, arg);
@@ -53,6 +61,7 @@ void Block::Reset(LocationDescriptor location_) noexcept {
     mcl::intrusive_list<IR::Inst> tmp = {};
     instructions.swap(tmp);
     inlined_inst.clear();
+    pooled_inst.clear();
     cond_failed.reset();
     location = location_;
     end_location = location_;

src/dynarmic/src/dynarmic/ir/basic_block.h

@@ -142,9 +142,11 @@ public:
     }
 
     /// "Hot cache" for small blocks so we don't call global allocator
-    boost::container::static_vector<Inst, 4096> inlined_inst = {};
+    boost::container::static_vector<Inst, 30> inlined_inst;
     /// List of instructions in this block.
     instruction_list_type instructions;
+    /// "Long/far" memory pool
+    boost::container::stable_vector<boost::container::static_vector<Inst, 32>> pooled_inst;
     /// Block to execute next if `cond` did not pass.
     std::optional<LocationDescriptor> cond_failed = {};
     /// Description of the starting location of this block
@@ -160,7 +162,7 @@ public:
     /// Number of cycles this block takes to execute.
     size_t cycle_count = 0;
 };
-//static_assert(sizeof(Block) == 4096);
+static_assert(sizeof(Block) == 4096);
 
 /// Returns a string representation of the contents of block. Intended for debugging.
 std::string DumpBlock(const IR::Block& block) noexcept;

src/video_core/CMakeLists.txt

@@ -20,6 +20,7 @@ add_library(video_core STATIC
     buffer_cache/buffer_cache.h
     buffer_cache/memory_tracker_base.h
     buffer_cache/usage_tracker.h
+    buffer_cache/virtual_range_cache.h
     buffer_cache/word_manager.h
     cache_types.h
     capture.h
@@ -166,6 +167,8 @@ add_library(video_core STATIC
     renderer_vulkan/vk_fence_manager.h
     renderer_vulkan/vk_graphics_pipeline.cpp
     renderer_vulkan/vk_graphics_pipeline.h
+    renderer_vulkan/vk_multi_range_buffer.cpp
+    renderer_vulkan/vk_multi_range_buffer.h
     renderer_vulkan/vk_master_semaphore.cpp
     renderer_vulkan/vk_master_semaphore.h
     renderer_vulkan/vk_pipeline_cache.cpp

src/video_core/buffer_cache/buffer_cache.h

@@ -112,6 +112,13 @@ void BufferCache<P>::TickFrame() {
     async_buffers_death_ring.clear();
 }
 
+template <class P>
+void BufferCache<P>::UnmapGPUMemory(size_t as_id, GPUVAddr gpu_addr, size_t size) {
+    if constexpr (requires { runtime.BindMultiRangeStorageBuffer(u64{}, bool{}); }) {
+        virtual_ranges.Unmap(as_id, gpu_addr, size);
+    }
+}
+
 template <class P>
 void BufferCache<P>::WriteMemory(DAddr device_addr, u64 size) {
     if (memory_tracker.IsRegionGpuModified(device_addr, size)) {
@@ -208,8 +215,8 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     BufferId buffer_b;
     do {
         channel_state->has_deleted_buffers = false;
-        buffer_a = FindBuffer(*cpu_src_address, static_cast<u32>(amount));
-        buffer_b = FindBuffer(*cpu_dest_address, static_cast<u32>(amount));
+        buffer_a = FindBuffer(*cpu_src_address, static_cast<u32>(amount), false);
+        buffer_b = FindBuffer(*cpu_dest_address, static_cast<u32>(amount), false);
     } while (channel_state->has_deleted_buffers);
     auto& src_buffer = slot_buffers[buffer_a];
     auto& dest_buffer = slot_buffers[buffer_b];
@@ -265,7 +272,7 @@ bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
     ClearDownload(*cpu_dst_address, size);
     gpu_modified_ranges.Subtract(*cpu_dst_address, size);
 
-    const BufferId buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size));
+    const BufferId buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size), false);
     Buffer& dest_buffer = slot_buffers[buffer];
     const u32 offset = dest_buffer.Offset(*cpu_dst_address);
     runtime.ClearBuffer(dest_buffer, offset, size, value);
@@ -287,7 +294,7 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_ad
 template <class P>
 std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(
     DAddr device_addr, u32 size, ObtainBufferSynchronize sync_info, ObtainBufferOperation post_op) {
-    const BufferId buffer_id = FindBuffer(device_addr, size);
+    const BufferId buffer_id = FindBuffer(device_addr, size, false);
     Buffer& buffer = slot_buffers[buffer_id];
 
     // synchronize op
@@ -998,11 +1005,85 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     channel_state->fast_bound_uniform_buffers[stage] &= ~(1u << binding_index);
 }
 
+template <class P>
+void BufferCache<P>::ResolveMultiRangeStorage(Binding& binding, bool is_written,
+                                              std::vector<MultiRangeSegment>& pool) {
+    binding.segment_first = 0;
+    binding.segment_count = 0;
+    if constexpr (requires { runtime.BindMultiRangeStorageBuffer(u64{}, bool{}); }) {
+        if (binding.gpu_addr == 0 || binding.size == 0) {
+            return;
+        }
+        if (is_written && !runtime.PrefersSparseSources()) {
+            return;
+        }
+        const VirtualSegments* found =
+            virtual_ranges.Query(*gpu_memory, binding.gpu_addr, binding.size);
+        if (!found || found->size() < 2) {
+            return;
+        }
+        const VirtualSegments segments = *found;
+        const u32 first = static_cast<u32>(pool.size());
+        const bool prefer_sparse = runtime.PrefersSparseSources();
+        for (const VirtualSegment& segment : segments) {
+            const BufferId buffer_id =
+                FindBuffer(segment.device_addr, segment.size, prefer_sparse);
+            if (!buffer_id) {
+                pool.resize(first);
+                return;
+            }
+            pool.push_back(MultiRangeSegment{
+                .buffer_id = buffer_id,
+                .device_addr = segment.device_addr,
+                .size = segment.size,
+            });
+        }
+        binding.segment_first = first;
+        binding.segment_count = static_cast<u32>(segments.size());
+    }
+}
+
+template <class P>
+bool BufferCache<P>::BindMultiRangeStorage(const Binding& binding, bool is_written,
+                                           std::span<const MultiRangeSegment> pool) {
+    if constexpr (requires { runtime.BindMultiRangeStorageBuffer(u64{}, bool{}); }) {
+        if (binding.segment_count < 2) {
+            return false;
+        }
+        if (binding.segment_first + binding.segment_count > pool.size()) {
+            return false;
+        }
+        const u64 key = (static_cast<u64>(gpu_memory->GetID()) << 48) ^ binding.gpu_addr;
+        runtime.ResetMultiRange();
+        for (u32 index = 0; index < binding.segment_count; ++index) {
+            const MultiRangeSegment& segment = pool[binding.segment_first + index];
+            Buffer& buffer = slot_buffers[segment.buffer_id];
+            TouchBuffer(buffer, segment.buffer_id);
+            if (SynchronizeBuffer(buffer, segment.device_addr, segment.size)) {
+                runtime.InvalidateMultiRange(key);
+            }
+            const u32 offset = buffer.Offset(segment.device_addr);
+            buffer.MarkUsage(offset, segment.size);
+            if (is_written) {
+                MarkWrittenBuffer(segment.buffer_id, segment.device_addr, segment.size);
+            }
+            runtime.PushMultiRangeSource(buffer, offset, segment.size);
+        }
+        return runtime.BindMultiRangeStorageBuffer(key, is_written);
+    } else {
+        return false;
+    }
+}
+
 template <class P>
 void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
     u32 binding_index = 0;
     ForEachEnabledBit(channel_state->enabled_storage_buffers[stage], [&](u32 index) {
         const Binding& binding = channel_state->storage_buffers[stage][index];
+        const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
+        if (BindMultiRangeStorage(binding, is_written, graphics_segments)) {
+            return;
+        }
         Buffer& buffer = slot_buffers[binding.buffer_id];
         TouchBuffer(buffer, binding.buffer_id);
         const u32 size = binding.size;
@@ -1010,7 +1091,6 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
 
         const u32 offset = buffer.Offset(binding.device_addr);
         buffer.MarkUsage(offset, size);
-        const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
 
         if (is_written) {
             MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size);
@@ -1139,6 +1219,11 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
     u32 binding_index = 0;
     ForEachEnabledBit(channel_state->enabled_compute_storage_buffers, [&](u32 index) {
         const Binding& binding = channel_state->compute_storage_buffers[index];
+        const bool is_written =
+            ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
+        if (BindMultiRangeStorage(binding, is_written, compute_segments)) {
+            return;
+        }
         Buffer& buffer = slot_buffers[binding.buffer_id];
         TouchBuffer(buffer, binding.buffer_id);
         const u32 size = binding.size;
@@ -1146,8 +1231,6 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
 
         const u32 offset = buffer.Offset(binding.device_addr);
         buffer.MarkUsage(offset, size);
-        const bool is_written =
-            ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
 
         if (is_written) {
             MarkWrittenBuffer(binding.buffer_id, binding.device_addr, size);
@@ -1193,6 +1276,7 @@ void BufferCache<P>::BindHostComputeTextureBuffers() {
 
 template <class P>
 void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) {
+    graphics_segments.clear();
     BufferOperations([&]() {
         if (is_indexed) {
             UpdateIndexBuffer();
@@ -1212,6 +1296,7 @@ void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) {
 
 template <class P>
 void BufferCache<P>::DoUpdateComputeBuffers() {
+    compute_segments.clear();
     BufferOperations([&]() {
         UpdateComputeUniformBuffers();
         UpdateComputeStorageBuffers();
@@ -1234,11 +1319,11 @@ void BufferCache<P>::UpdateIndexBuffer() {
         auto inline_index_size = static_cast<u32>(draw_state.inline_index_draw_indexes.size());
         u32 buffer_size = Common::AlignUp(inline_index_size, CACHING_PAGESIZE);
         if (inline_buffer_id == NULL_BUFFER_ID) [[unlikely]] {
-            inline_buffer_id = CreateBuffer(0, buffer_size);
+            inline_buffer_id = CreateBuffer(0, buffer_size, false);
         }
         if (slot_buffers[inline_buffer_id].SizeBytes() < buffer_size) [[unlikely]] {
             slot_buffers.erase(inline_buffer_id);
-            inline_buffer_id = CreateBuffer(0, buffer_size);
+            inline_buffer_id = CreateBuffer(0, buffer_size, false);
         }
         channel_state->index_buffer = Binding{
             .device_addr = 0,
@@ -1261,7 +1346,7 @@ void BufferCache<P>::UpdateIndexBuffer() {
     channel_state->index_buffer = Binding{
         .device_addr = *device_addr,
         .size = size,
-        .buffer_id = FindBuffer(*device_addr, size),
+        .buffer_id = FindBuffer(*device_addr, size, false),
     };
 }
 
@@ -1298,7 +1383,7 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) {
     if (!gpu_memory->IsWithinGPUAddressRange(gpu_addr_end) || size >= 64_MiB) {
         size = static_cast<u32>(gpu_memory->MaxContinuousRange(gpu_addr_begin, size));
     }
-    const BufferId buffer_id = FindBuffer(*device_addr, size);
+    const BufferId buffer_id = FindBuffer(*device_addr, size, false);
     const Binding binding{
         .device_addr = *device_addr,
         .size = size,
@@ -1319,7 +1404,7 @@ void BufferCache<P>::UpdateDrawIndirect() {
         binding = Binding{
             .device_addr = *device_addr,
             .size = static_cast<u32>(size),
-            .buffer_id = FindBuffer(*device_addr, static_cast<u32>(size)),
+            .buffer_id = FindBuffer(*device_addr, static_cast<u32>(size), false),
         };
     };
     if (current_draw_indirect->include_count) {
@@ -1343,7 +1428,7 @@ void BufferCache<P>::UpdateUniformBuffers(size_t stage) {
             channel_state->dirty_uniform_buffers[stage] |= 1U << index;
         }
         // Resolve buffer
-        binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        binding.buffer_id = FindBuffer(binding.device_addr, binding.size, false);
     });
 }
 
@@ -1352,8 +1437,10 @@ void BufferCache<P>::UpdateStorageBuffers(size_t stage) {
     ForEachEnabledBit(channel_state->enabled_storage_buffers[stage], [&](u32 index) {
         // Resolve buffer
         Binding& binding = channel_state->storage_buffers[stage][index];
-        const BufferId buffer_id = FindBuffer(binding.device_addr, binding.size);
+        const BufferId buffer_id = FindBuffer(binding.device_addr, binding.size, false);
         binding.buffer_id = buffer_id;
+        const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
+        ResolveMultiRangeStorage(binding, is_written, graphics_segments);
     });
 }
 
@@ -1361,7 +1448,7 @@ template <class P>
 void BufferCache<P>::UpdateTextureBuffers(size_t stage) {
     ForEachEnabledBit(channel_state->enabled_texture_buffers[stage], [&](u32 index) {
         Binding& binding = channel_state->texture_buffers[stage][index];
-        binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        binding.buffer_id = FindBuffer(binding.device_addr, binding.size, false);
     });
 }
 
@@ -1385,7 +1472,7 @@ void BufferCache<P>::UpdateTransformFeedbackBuffer(u32 index) {
         channel_state->transform_feedback_buffers[index] = NULL_BINDING;
         return;
     }
-    const BufferId buffer_id = FindBuffer(*device_addr, size);
+    const BufferId buffer_id = FindBuffer(*device_addr, size, false);
     channel_state->transform_feedback_buffers[index] = Binding{
         .device_addr = *device_addr,
         .size = size,
@@ -1407,7 +1494,7 @@ void BufferCache<P>::UpdateComputeUniformBuffers() {
                 binding.size = cbuf.size;
             }
         }
-        binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        binding.buffer_id = FindBuffer(binding.device_addr, binding.size, false);
     });
 }
 
@@ -1416,7 +1503,10 @@ void BufferCache<P>::UpdateComputeStorageBuffers() {
     ForEachEnabledBit(channel_state->enabled_compute_storage_buffers, [&](u32 index) {
         // Resolve buffer
         Binding& binding = channel_state->compute_storage_buffers[index];
-        binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        binding.buffer_id = FindBuffer(binding.device_addr, binding.size, false);
+        const bool is_written =
+            ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
+        ResolveMultiRangeStorage(binding, is_written, compute_segments);
     });
 }
 
@@ -1424,7 +1514,7 @@ template <class P>
 void BufferCache<P>::UpdateComputeTextureBuffers() {
     ForEachEnabledBit(channel_state->enabled_compute_texture_buffers, [&](u32 index) {
         Binding& binding = channel_state->compute_texture_buffers[index];
-        binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        binding.buffer_id = FindBuffer(binding.device_addr, binding.size, false);
     });
 }
 
@@ -1440,7 +1530,7 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u3
 }
 
 template <class P>
-BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size) {
+BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size, bool sparse_compatible) {
     if (device_addr == 0) {
         return NULL_BUFFER_ID;
     }
@@ -1450,10 +1540,18 @@ BufferId BufferCache<P>::FindBuffer(DAddr device_addr, u32 size) {
         Buffer& buffer = slot_buffers[buffer_id];
         WaitForGpuFenceIfNeeded(buffer);
         if (buffer.IsInBounds(device_addr, size)) {
-            return buffer_id;
+            bool usable = true;
+            if constexpr (requires { buffer.IsSparseCompatible(); }) {
+                if (sparse_compatible && !buffer.IsSparseCompatible()) {
+                    usable = false;
+                }
+            }
+            if (usable) {
+                return buffer_id;
+            }
         }
     }
-    return CreateBuffer(device_addr, size);
+    return CreateBuffer(device_addr, size, sparse_compatible);
 }
 
 template <class P>
@@ -1575,13 +1673,15 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
 }
 
 template <class P>
-BufferId BufferCache<P>::CreateBuffer(DAddr device_addr, u32 wanted_size) {
+BufferId BufferCache<P>::CreateBuffer(DAddr device_addr, u32 wanted_size,
+                                      bool sparse_compatible) {
     DAddr device_addr_end = Common::AlignUp(device_addr + wanted_size, CACHING_PAGESIZE);
     device_addr = Common::AlignDown(device_addr, CACHING_PAGESIZE);
     wanted_size = static_cast<u32>(device_addr_end - device_addr);
     const OverlapResult overlap = ResolveOverlaps(device_addr, wanted_size);
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
-    const BufferId new_buffer_id = slot_buffers.insert(runtime, overlap.begin, size);
+    const BufferId new_buffer_id =
+        slot_buffers.insert(runtime, overlap.begin, size, sparse_compatible);
     auto& new_buffer = slot_buffers[new_buffer_id];
     const size_t size_bytes = new_buffer.SizeBytes();
     runtime.ClearBuffer(new_buffer, 0, size_bytes, 0);
@@ -1745,7 +1845,7 @@ void BufferCache<P>::InlineMemoryImplementation(DAddr dest_address, size_t copy_
     ClearDownload(dest_address, copy_size);
     gpu_modified_ranges.Subtract(dest_address, copy_size);
 
-    BufferId buffer_id = FindBuffer(dest_address, static_cast<u32>(copy_size));
+    BufferId buffer_id = FindBuffer(dest_address, static_cast<u32>(copy_size), false);
     auto& buffer = slot_buffers[buffer_id];
     SynchronizeBuffer(buffer, dest_address, static_cast<u32>(copy_size));
 
@@ -1831,6 +1931,9 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, DAddr device_addr, u64
 
 template <class P>
 void BufferCache<P>::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
+    if constexpr (requires { runtime.OnBufferDeleted(slot_buffers[buffer_id]); }) {
+        runtime.OnBufferDeleted(slot_buffers[buffer_id]);
+    }
     bool dirty_index{false};
     boost::container::small_vector<u64, NUM_VERTEX_BUFFERS> dirty_vertex_buffers;
     const auto scalar_replace = [buffer_id](Binding& binding) {
@@ -1934,9 +2037,14 @@ Binding BufferCache<P>::StorageBufferBinding(GPUVAddr ssbo_addr, u32 cbuf_index,
     // The end address used for size calculation does not need to be aligned
     const DAddr cpu_end = Common::AlignUp(*device_addr + size, Core::DEVICE_PAGESIZE);
 
+    u32 binding_size = static_cast<u32>(cpu_end - *aligned_device_addr);
+    if (is_written) {
+        binding_size = aligned_size;
+    }
     const Binding binding{
         .device_addr = *aligned_device_addr,
-        .size = is_written ? aligned_size : static_cast<u32>(cpu_end - *aligned_device_addr),
+        .gpu_addr = aligned_gpu_addr,
+        .size = binding_size,
         .buffer_id = BufferId{},
     };
     return binding;

src/video_core/buffer_cache/buffer_cache_base.h

@@ -29,6 +29,7 @@
 #include "common/settings.h"
 #include "common/slot_vector.h"
 #include "video_core/buffer_cache/buffer_base.h"
+#include "video_core/buffer_cache/virtual_range_cache.h"
 #include "video_core/control/channel_state_cache.h"
 #include "video_core/delayed_destruction_ring.h"
 #include "video_core/dirty_flags.h"
@@ -81,8 +82,17 @@ static constexpr u32 DEFAULT_SKIP_CACHE_SIZE = static_cast<u32>(4_KiB);
 
 struct Binding {
     DAddr device_addr{};
+    GPUVAddr gpu_addr{};
     u32 size{};
     BufferId buffer_id;
+    u32 segment_first{};
+    u32 segment_count{};
+};
+
+struct MultiRangeSegment {
+    BufferId buffer_id;
+    DAddr device_addr{};
+    u32 size{};
 };
 
 struct TextureBufferBinding : Binding {
@@ -215,6 +225,14 @@ public:
 
     void TickFrame();
 
+    bool BindMultiRangeStorage(const Binding& binding, bool is_written,
+                               std::span<const MultiRangeSegment> pool);
+
+    void ResolveMultiRangeStorage(Binding& binding, bool is_written,
+                                  std::vector<MultiRangeSegment>& pool);
+
+    void UnmapGPUMemory(size_t as_id, GPUVAddr gpu_addr, size_t size);
+
     void WriteMemory(DAddr device_addr, u64 size);
 
     void CachedWriteMemory(DAddr device_addr, u64 size);
@@ -414,7 +432,7 @@ private:
 
     void MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u32 size);
 
-    [[nodiscard]] BufferId FindBuffer(DAddr device_addr, u32 size);
+    [[nodiscard]] BufferId FindBuffer(DAddr device_addr, u32 size, bool sparse_compatible);
 
     void WaitForGpuFenceIfNeeded(Buffer& buffer);
 
@@ -422,7 +440,8 @@ private:
 
     void JoinOverlap(BufferId new_buffer_id, BufferId overlap_id, bool accumulate_stream_score);
 
-    [[nodiscard]] BufferId CreateBuffer(DAddr device_addr, u32 wanted_size);
+    [[nodiscard]] BufferId CreateBuffer(DAddr device_addr, u32 wanted_size,
+                                        bool sparse_compatible);
 
     void Register(BufferId buffer_id);
 
@@ -513,6 +532,9 @@ private:
         using TickType = u64;
     };
     Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
+    VirtualRangeCache virtual_ranges;
+    std::vector<MultiRangeSegment> graphics_segments;
+    std::vector<MultiRangeSegment> compute_segments;
     u64 frame_tick = 0;
     u64 total_used_memory = 0;
     u64 minimum_memory = 0;

src/video_core/buffer_cache/virtual_range_cache.h

@@ -0,0 +1,173 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <atomic>
+#include <limits>
+#include <mutex>
+#include <optional>
+#include <vector>
+
+#include <boost/container/small_vector.hpp>
+
+#include "common/common_types.h"
+#include "common/container/unordered_map.h"
+#include "video_core/memory_manager.h"
+
+namespace VideoCommon {
+
+struct VirtualSegment {
+    GPUVAddr gpu_addr;
+    DAddr device_addr;
+    u32 size;
+};
+
+using VirtualSegments = boost::container::small_vector<VirtualSegment, 8>;
+
+class VirtualRangeCache {
+public:
+    static constexpr size_t MAX_ENTRIES = 8192;
+    static constexpr size_t MAX_DEFERRED = 4096;
+
+    const VirtualSegments* Query(Tegra::MemoryManager& memory, GPUVAddr gpu_addr, u32 size) {
+        if (has_deferred.load(std::memory_order_acquire)) {
+            ApplyDeferred();
+        }
+        if (entries.size() > MAX_ENTRIES) {
+            entries.clear();
+        }
+        const size_t as_id = memory.GetID();
+        const u64 key = MakeKey(as_id, gpu_addr);
+        const auto it = entries.find(key);
+        if (it != entries.end() && it->second.as_id == as_id &&
+            it->second.gpu_addr == gpu_addr && it->second.size == size) {
+            return &it->second.segments;
+        }
+        Entry entry{};
+        entry.as_id = as_id;
+        entry.gpu_addr = gpu_addr;
+        entry.size = size;
+        const auto ranges = memory.GetSubmappedRange(gpu_addr, size);
+        GPUVAddr expected = gpu_addr;
+        bool contiguous = true;
+        for (const auto& [range_addr, range_size] : ranges) {
+            if (range_addr != expected || range_size == 0) {
+                contiguous = false;
+                break;
+            }
+            const std::optional<DAddr> device_addr = memory.GpuToCpuAddress(range_addr);
+            if (!device_addr || *device_addr == 0) {
+                contiguous = false;
+                break;
+            }
+            if (range_size > static_cast<size_t>((std::numeric_limits<u32>::max)())) {
+                contiguous = false;
+                break;
+            }
+            entry.segments.push_back(VirtualSegment{
+                .gpu_addr = range_addr,
+                .device_addr = *device_addr,
+                .size = static_cast<u32>(range_size),
+            });
+            expected += range_size;
+        }
+        if (!contiguous || expected != gpu_addr + size) {
+            entry.segments.clear();
+        }
+        const auto result = entries.insert_or_assign(key, std::move(entry));
+        return &result.first->second.segments;
+    }
+
+    void Unmap(size_t as_id, GPUVAddr gpu_addr, u64 size) {
+        if (size == 0) {
+            return;
+        }
+        {
+            std::scoped_lock lock{deferred_mutex};
+            if (!deferred.empty()) {
+                DeferredUnmap& last = deferred.back();
+                if (last.as_id == as_id && last.gpu_addr + last.size == gpu_addr) {
+                    last.size += size;
+                    has_deferred.store(true, std::memory_order_release);
+                    return;
+                }
+            }
+            if (deferred.size() >= MAX_DEFERRED) {
+                deferred.clear();
+                deferred_overflow = true;
+            } else {
+                deferred.push_back(DeferredUnmap{
+                    .as_id = as_id,
+                    .gpu_addr = gpu_addr,
+                    .size = size,
+                });
+            }
+        }
+        has_deferred.store(true, std::memory_order_release);
+    }
+
+private:
+    struct Entry {
+        VirtualSegments segments;
+        size_t as_id{};
+        GPUVAddr gpu_addr{};
+        u32 size{};
+    };
+
+    struct DeferredUnmap {
+        size_t as_id;
+        GPUVAddr gpu_addr;
+        u64 size;
+    };
+
+    static u64 MakeKey(size_t as_id, GPUVAddr gpu_addr) {
+        return (static_cast<u64>(as_id) << 48) ^ gpu_addr;
+    }
+
+    void ApplyDeferred() {
+        std::vector<DeferredUnmap> pending;
+        bool overflow = false;
+        {
+            std::scoped_lock lock{deferred_mutex};
+            has_deferred.store(false, std::memory_order_release);
+            pending.swap(deferred);
+            overflow = deferred_overflow;
+            deferred_overflow = false;
+        }
+        if (overflow) {
+            entries.clear();
+            return;
+        }
+        if (pending.empty() || entries.empty()) {
+            return;
+        }
+        for (auto it = entries.begin(); it != entries.end();) {
+            const Entry& entry = it->second;
+            const GPUVAddr entry_end = entry.gpu_addr + entry.size;
+            bool overlaps = false;
+            for (const DeferredUnmap& unmap : pending) {
+                if (unmap.as_id != entry.as_id) {
+                    continue;
+                }
+                if (entry.gpu_addr < unmap.gpu_addr + unmap.size && unmap.gpu_addr < entry_end) {
+                    overlaps = true;
+                    break;
+                }
+            }
+            if (overlaps) {
+                it = entries.erase(it);
+            } else {
+                ++it;
+            }
+        }
+    }
+
+    ::Common::unordered_map<u64, Entry> entries;
+    std::vector<DeferredUnmap> deferred;
+    std::mutex deferred_mutex;
+    std::atomic<bool> has_deferred{false};
+    bool deferred_overflow{};
+};
+
+} // namespace VideoCommon

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -52,7 +52,7 @@ constexpr std::array PROGRAM_LUT{
 Buffer::Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params)
     : VideoCommon::BufferBase(null_params) {}
 
-Buffer::Buffer(BufferCacheRuntime& runtime, DAddr cpu_addr_, u64 size_bytes_)
+Buffer::Buffer(BufferCacheRuntime& runtime, DAddr cpu_addr_, u64 size_bytes_, bool)
     : VideoCommon::BufferBase(cpu_addr_, size_bytes_) {
     buffer.Create();
     if (runtime.device.HasDebuggingToolAttached()) {

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -23,7 +23,8 @@ class BufferCacheRuntime;
 
 class Buffer : public VideoCommon::BufferBase {
 public:
-    explicit Buffer(BufferCacheRuntime&, DAddr cpu_addr, u64 size_bytes);
+    explicit Buffer(BufferCacheRuntime&, DAddr cpu_addr, u64 size_bytes,
+                    bool sparse_compatible);
     explicit Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams);
 
     void ImmediateUpload(size_t offset, std::span<const u8> data) noexcept;

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -56,7 +56,8 @@ size_t BytesPerIndex(VkIndexType index_type) {
     }
 }
 
-vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allocator, u64 size) {
+vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allocator, u64 size,
+                        VkDeviceSize sparse_alignment) {
     VkBufferUsageFlags flags =
         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 |
@@ -82,6 +83,9 @@ vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allo
         .queueFamilyIndexCount = 0,
         .pQueueFamilyIndices = nullptr,
     };
+    if (sparse_alignment > 1) {
+        return memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal, sparse_alignment);
+    }
     return memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::DeviceLocal);
 }
 } // Anonymous namespace
@@ -99,10 +103,14 @@ Buffer::Buffer(BufferCacheRuntime& runtime, VideoCommon::NullBufferParams null_p
     }
 }
 
-Buffer::Buffer(BufferCacheRuntime& runtime, DAddr cpu_addr_, u64 size_bytes_)
+Buffer::Buffer(BufferCacheRuntime& runtime, DAddr cpu_addr_, u64 size_bytes_,
+               bool sparse_compatible_)
     : VideoCommon::BufferBase(cpu_addr_, size_bytes_), device{&runtime.device},
       scheduler{&runtime.scheduler},
-      buffer{CreateBuffer(*device, runtime.memory_allocator, SizeBytes())}, tracker{SizeBytes()} {
+      buffer{CreateBuffer(*device, runtime.memory_allocator, SizeBytes(),
+                          runtime.SparseAlignmentFor(sparse_compatible_))},
+      tracker{SizeBytes()} {
+    sparse_compatible = sparse_compatible_;
     if (runtime.device.HasDebuggingToolAttached()) {
         buffer.SetObjectNameEXT(fmt::format("Buffer {:#x}", CpuAddr()).c_str());
     }
@@ -348,7 +356,8 @@ BufferCacheRuntime::BufferCacheRuntime(const Device& device_, MemoryAllocator& m
     : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
       staging_pool{staging_pool_}, guest_descriptor_queue{guest_descriptor_queue_},
       quad_index_pass(device, scheduler, descriptor_pool, staging_pool,
-                      compute_pass_descriptor_queue) {
+                      compute_pass_descriptor_queue),
+      multi_range_buffers(device_) {
     const VkDriverIdKHR driver_id = device.GetDriverID();
     limit_dynamic_storage_buffers = driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY ||
                                     driver_id == VK_DRIVER_ID_ARM_PROPRIETARY;
@@ -536,6 +545,37 @@ void BufferCacheRuntime::ClearBuffer(VkBuffer dest_buffer, u32 offset, size_t si
     });
 }
 
+bool BufferCacheRuntime::BindMultiRangeStorageBuffer(u64 key, bool is_written) {
+    if (multi_range_sources.empty() || multi_range_total == 0) {
+        return false;
+    }
+    const MultiRangeRef ref = multi_range_buffers.Get(device, scheduler, memory_allocator, key,
+                                                     multi_range_sources, multi_range_total);
+    if (ref.handle == VK_NULL_HANDLE) {
+        return false;
+    }
+    if (is_written && !ref.sparse) {
+        return false;
+    }
+    if (ref.needs_gather) {
+        PreCopyBarrier();
+        VkDeviceSize dst_offset = 0;
+        for (const MultiRangeSource& source : multi_range_sources) {
+            const std::array<VideoCommon::BufferCopy, 1> copy{VideoCommon::BufferCopy{
+                .src_offset = u64(source.offset),
+                .dst_offset = u64(dst_offset),
+                .size = size_t(source.size),
+            }};
+            CopyBuffer(ref.handle, source.handle, copy, false);
+            dst_offset += source.size;
+        }
+        PostCopyBarrier();
+        multi_range_buffers.MarkGathered(key);
+    }
+    guest_descriptor_queue.AddBuffer(ref.handle, ref.address, 0, ref.size);
+    return true;
+}
+
 void BufferCacheRuntime::BindIndexBuffer(PrimitiveTopology topology, IndexFormat index_format,
                                          u32 base_vertex, u32 num_indices, VkBuffer buffer,
                                          u32 offset, [[maybe_unused]] u32 size) {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -8,11 +8,14 @@
 
 #include <limits>
 
+#include <boost/container/small_vector.hpp>
+
 #include "video_core/buffer_cache/buffer_cache_base.h"
 #include "video_core/buffer_cache/memory_tracker_base.h"
 #include "video_core/buffer_cache/usage_tracker.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
+#include "video_core/renderer_vulkan/vk_multi_range_buffer.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
 #include "video_core/renderer_vulkan/vk_update_descriptor.h"
 #include "video_core/surface.h"
@@ -31,7 +34,8 @@ class BufferCacheRuntime;
 class Buffer : public VideoCommon::BufferBase {
 public:
     explicit Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params);
-    explicit Buffer(BufferCacheRuntime& runtime, VAddr cpu_addr_, u64 size_bytes_);
+    explicit Buffer(BufferCacheRuntime& runtime, VAddr cpu_addr_, u64 size_bytes_,
+                    bool sparse_compatible_);
 
     [[nodiscard]] VkBufferView View(u32 offset, u32 size, VideoCore::Surface::PixelFormat format);
 
@@ -43,6 +47,14 @@ public:
         return device_address;
     }
 
+    [[nodiscard]] bool IsSparseCompatible() const noexcept {
+        return sparse_compatible;
+    }
+
+    [[nodiscard]] vk::MemoryLocation Location() const noexcept {
+        return buffer.Location();
+    }
+
     [[nodiscard]] bool IsRegionUsed(u64 offset, u64 size) const noexcept {
         return tracker.IsUsed(offset, size);
     }
@@ -77,6 +89,7 @@ private:
     VkDeviceAddress device_address{};
     u64 last_usage_tick{};
     bool is_null{};
+    bool sparse_compatible{};
 };
 
 class QuadArrayIndexBuffer;
@@ -125,7 +138,7 @@ public:
 
     void PreCopyBarrier();
 
-    void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
+    void CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
                     std::span<const VideoCommon::BufferCopy> copies, bool barrier,
                     bool can_reorder_upload = false);
 
@@ -155,6 +168,46 @@ public:
         return ref.mapped_span;
     }
 
+    [[nodiscard]] VkDeviceSize SparseAlignmentFor(bool sparse_compatible) const noexcept {
+        if (!sparse_compatible || !multi_range_buffers.use_sparse) {
+            return 0;
+        }
+        return multi_range_buffers.block_size;
+    }
+
+    [[nodiscard]] bool PrefersSparseSources() const noexcept {
+        return multi_range_buffers.use_sparse;
+    }
+
+    void ResetMultiRange() noexcept {
+        multi_range_sources.clear();
+        multi_range_total = 0;
+    }
+
+    void PushMultiRangeSource(const Buffer& buffer, u32 offset, u32 size) {
+        const vk::MemoryLocation location = buffer.Location();
+        multi_range_sources.push_back(MultiRangeSource{
+            .handle = buffer.Handle(),
+            .memory = location.memory,
+            .memory_offset = location.offset,
+            .offset = offset,
+            .size = size,
+            .write_tick = buffer.getWriteTick(),
+            .memory_type = location.memory_type,
+        });
+        multi_range_total += size;
+    }
+
+    bool BindMultiRangeStorageBuffer(u64 key, bool is_written);
+
+    void InvalidateMultiRange(u64 key) {
+        multi_range_buffers.Invalidate(key);
+    }
+
+    void OnBufferDeleted(const Buffer& buffer) {
+        multi_range_buffers.DropOwner(scheduler, buffer.Handle());
+    }
+
     void BindUniformBuffer(const Buffer& buffer, u32 offset, u32 size) {
         BindBuffer(buffer, offset, size);
     }
@@ -208,6 +261,10 @@ private:
     std::unique_ptr<Uint8Pass> uint8_pass;
     QuadIndexedPass quad_index_pass;
 
+    MultiRangeBufferCache multi_range_buffers;
+    boost::container::small_vector<MultiRangeSource, 16> multi_range_sources;
+    VkDeviceSize multi_range_total{};
+
     bool limit_dynamic_storage_buffers = false;
     u32 max_dynamic_storage_buffers = (std::numeric_limits<u32>::max)();
 };

src/video_core/renderer_vulkan/vk_multi_range_buffer.cpp

@@ -0,0 +1,338 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <algorithm>
+#include <mutex>
+#include <utility>
+
+#include "video_core/renderer_vulkan/vk_multi_range_buffer.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/vulkan_common/vulkan_device.h"
+
+namespace Vulkan {
+
+MultiRangeBufferCache::MultiRangeBufferCache(const Device& device) {
+    sparse_usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+                   VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+    if (device.IsBufferDeviceAddressSupported()) {
+        sparse_usage |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+    }
+    if (!device.IsSparseBindingSupported()) {
+        return;
+    }
+    u32 memory_type_bits = 0;
+    const VkDeviceSize queried = QueryBlockSize(device, memory_type_bits);
+    if (queried == 0 || memory_type_bits == 0) {
+        return;
+    }
+    block_size = queried;
+    sparse_memory_type_bits = memory_type_bits;
+    use_sparse = true;
+}
+
+VkDeviceSize MultiRangeBufferCache::QueryBlockSize(const Device& device,
+                                                   u32& memory_type_bits) const {
+    const VkDevice logical = *device.GetLogical();
+    const auto& dld = device.GetDispatchLoader();
+    const VkBufferCreateInfo probe_ci{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = VK_BUFFER_CREATE_SPARSE_BINDING_BIT | VK_BUFFER_CREATE_SPARSE_ALIASED_BIT,
+        .size = DEFAULT_BLOCK_SIZE,
+        .usage = sparse_usage,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    };
+    VkBuffer probe{};
+    if (dld.vkCreateBuffer(logical, &probe_ci, nullptr, &probe) != VK_SUCCESS) {
+        return 0;
+    }
+    const SparseBuffer owned{probe, logical, dld};
+    const VkBufferMemoryRequirementsInfo2 reqs_info{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
+        .pNext = nullptr,
+        .buffer = probe,
+    };
+    VkMemoryRequirements2 reqs2{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
+        .pNext = nullptr,
+        .memoryRequirements = {},
+    };
+    dld.vkGetBufferMemoryRequirements2(logical, &reqs_info, &reqs2);
+    memory_type_bits = reqs2.memoryRequirements.memoryTypeBits;
+    return reqs2.memoryRequirements.alignment;
+}
+
+u64 MultiRangeBufferCache::HashSources(std::span<const MultiRangeSource> sources) const {
+    u64 hash = 0xcbf29ce484222325ULL;
+    const auto mix = [&hash](u64 value) {
+        hash ^= value;
+        hash *= 0x100000001b3ULL;
+    };
+    for (const MultiRangeSource& source : sources) {
+        mix(u64(source.handle));
+        mix(u64(source.offset));
+        mix(u64(source.size));
+    }
+    return hash;
+}
+
+u64 MultiRangeBufferCache::HashContent(std::span<const MultiRangeSource> sources) const {
+    u64 hash = 0xcbf29ce484222325ULL;
+    for (const MultiRangeSource& source : sources) {
+        hash ^= source.write_tick;
+        hash *= 0x100000001b3ULL;
+    }
+    return hash;
+}
+
+bool MultiRangeBufferCache::CanBindSparse(std::span<const MultiRangeSource> sources) const {
+    return use_sparse &&
+           std::none_of(sources.begin(), sources.end(),
+                        [block = block_size, bits = sparse_memory_type_bits](auto const& e) {
+                            const VkDeviceSize memory_offset = e.memory_offset + e.offset;
+                            return e.memory == VK_NULL_HANDLE || e.memory_type >= 32 ||
+                                   ((bits >> e.memory_type) & 1) == 0 ||
+                                   (memory_offset % block) != 0 || (e.size % block) != 0;
+                        });
+}
+
+SparseBuffer MultiRangeBufferCache::CreateSparse(const Device& device, Scheduler& scheduler,
+                                                 std::span<const MultiRangeSource> sources,
+                                                 VkDeviceSize total) {
+    const VkDevice logical = *device.GetLogical();
+    const auto& dld = device.GetDispatchLoader();
+    const VkBufferCreateInfo buffer_ci{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = VK_BUFFER_CREATE_SPARSE_BINDING_BIT | VK_BUFFER_CREATE_SPARSE_ALIASED_BIT,
+        .size = total,
+        .usage = sparse_usage,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    };
+    VkBuffer raw{};
+    if (dld.vkCreateBuffer(logical, &buffer_ci, nullptr, &raw) != VK_SUCCESS) {
+        return SparseBuffer{};
+    }
+    SparseBuffer handle{raw, logical, dld};
+    std::vector<VkSparseMemoryBind> binds;
+    binds.reserve(sources.size());
+    VkDeviceSize resource_offset = 0;
+    for (const MultiRangeSource& source : sources) {
+        binds.push_back(VkSparseMemoryBind{
+            .resourceOffset = resource_offset,
+            .size = source.size,
+            .memory = source.memory,
+            .memoryOffset = source.memory_offset + source.offset,
+            .flags = 0,
+        });
+        resource_offset += source.size;
+    }
+    const VkSparseBufferMemoryBindInfo buffer_bind{
+        .buffer = raw,
+        .bindCount = static_cast<u32>(binds.size()),
+        .pBinds = binds.data(),
+    };
+    const VkBindSparseInfo bind_info{
+        .sType = VK_STRUCTURE_TYPE_BIND_SPARSE_INFO,
+        .pNext = nullptr,
+        .waitSemaphoreCount = 0,
+        .pWaitSemaphores = nullptr,
+        .bufferBindCount = 1,
+        .pBufferBinds = &buffer_bind,
+        .imageOpaqueBindCount = 0,
+        .pImageOpaqueBinds = nullptr,
+        .imageBindCount = 0,
+        .pImageBinds = nullptr,
+        .signalSemaphoreCount = 0,
+        .pSignalSemaphores = nullptr,
+    };
+    const VkFenceCreateInfo fence_ci{
+        .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+    };
+    vk::Fence fence = device.GetLogical().CreateFence(fence_ci);
+    VkResult bind_result = VK_ERROR_UNKNOWN;
+    {
+        std::scoped_lock lock{scheduler.submit_mutex};
+        bind_result = device.GetGraphicsQueue().BindSparse(bind_info, *fence);
+    }
+    if (bind_result != VK_SUCCESS) {
+        return SparseBuffer{};
+    }
+    fence.Wait();
+    return handle;
+}
+
+void MultiRangeBufferCache::RetireEntry(Scheduler& scheduler, Entry& entry) {
+    if (!entry.sparse_handle && !entry.gathered) {
+        return;
+    }
+    if (retired.size() == retired.capacity()) {
+        DrainRetired(scheduler);
+    }
+    if (retired.size() == retired.capacity()) {
+        u64 oldest = retired.front().tick;
+        for (const Retired& item : retired) {
+            if (item.tick < oldest) {
+                oldest = item.tick;
+            }
+        }
+        scheduler.Wait(oldest);
+        DrainRetired(scheduler);
+    }
+    retired.push_back(Retired{
+        .handle = std::move(entry.sparse_handle),
+        .gathered = std::move(entry.gathered),
+        .tick = scheduler.CurrentTick(),
+    });
+}
+
+void MultiRangeBufferCache::DrainRetired(Scheduler& scheduler) {
+    size_t index = 0;
+    while (index < retired.size()) {
+        if (scheduler.IsFree(retired[index].tick)) {
+            if (index + 1 != retired.size()) {
+                retired[index] = std::move(retired.back());
+            }
+            retired.pop_back();
+        } else {
+            ++index;
+        }
+    }
+}
+
+MultiRangeRef MultiRangeBufferCache::Get(const Device& device, Scheduler& scheduler,
+                                         MemoryAllocator& memory_allocator, u64 key,
+                                         std::span<const MultiRangeSource> sources,
+                                         VkDeviceSize total) {
+    if (sources.empty() || total == 0) {
+        return MultiRangeRef{};
+    }
+    if (!retired.empty()) {
+        DrainRetired(scheduler);
+    }
+    const u64 geometry = HashSources(sources);
+    const u64 content = HashContent(sources);
+    const auto it = entries.find(key);
+    if (it != entries.end() && it->second.geometry == geometry && it->second.size == total) {
+        Entry& entry = it->second;
+        if (entry.content != content) {
+            entry.content = content;
+            entry.dirty = true;
+        }
+        MultiRangeRef ref{
+            .handle = *entry.sparse_handle,
+            .address = entry.address,
+            .size = entry.size,
+            .sparse = true,
+            .needs_gather = false,
+        };
+        if (!entry.sparse_handle) {
+            ref.handle = *entry.gathered;
+            ref.sparse = false;
+            ref.needs_gather = entry.dirty;
+        }
+        return ref;
+    }
+    if (it != entries.end()) {
+        RetireEntry(scheduler, it->second);
+        entries.erase(it);
+    }
+
+    Entry entry{};
+    entry.geometry = geometry;
+    entry.content = content;
+    entry.size = total;
+    if (CanBindSparse(sources)) {
+        entry.sparse_handle = CreateSparse(device, scheduler, sources, total);
+        if (entry.sparse_handle) {
+            entry.owners.reserve(sources.size());
+            for (const MultiRangeSource& source : sources) {
+                entry.owners.push_back(source.handle);
+            }
+        }
+    }
+    if (!entry.sparse_handle) {
+        VkBufferUsageFlags flags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
+                                   VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+                                   VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+        if (device.IsBufferDeviceAddressSupported()) {
+            flags |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+        }
+        const VkBufferCreateInfo gather_ci{
+            .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .size = total,
+            .usage = flags,
+            .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+            .queueFamilyIndexCount = 0,
+            .pQueueFamilyIndices = nullptr,
+        };
+        entry.gathered = memory_allocator.CreateBuffer(gather_ci, MemoryUsage::DeviceLocal);
+        entry.dirty = true;
+    }
+    if (device.IsBufferDeviceAddressSupported()) {
+        VkBuffer address_handle = *entry.sparse_handle;
+        if (!entry.sparse_handle) {
+            address_handle = *entry.gathered;
+        }
+        entry.address = device.GetLogical().GetBufferDeviceAddress(address_handle);
+    }
+
+    MultiRangeRef ref{
+        .handle = *entry.sparse_handle,
+        .address = entry.address,
+        .size = entry.size,
+        .sparse = true,
+        .needs_gather = false,
+    };
+    if (!entry.sparse_handle) {
+        ref.handle = *entry.gathered;
+        ref.sparse = false;
+        ref.needs_gather = true;
+    }
+    entries.emplace(key, std::move(entry));
+    return ref;
+}
+
+void MultiRangeBufferCache::MarkGathered(u64 key) {
+    if (auto const it = entries.find(key); it != entries.end()) {
+        it->second.dirty = false;
+    }
+}
+
+void MultiRangeBufferCache::DropOwner(Scheduler& scheduler, VkBuffer owner) {
+    if (owner == VK_NULL_HANDLE) {
+        return;
+    }
+    for (auto it = entries.begin(); it != entries.end();) {
+        Entry& entry = it->second;
+        bool owned = false;
+        for (const VkBuffer handle : entry.owners) {
+            if (handle == owner) {
+                owned = true;
+                break;
+            }
+        }
+        if (!owned) {
+            ++it;
+            continue;
+        }
+        RetireEntry(scheduler, entry);
+        it = entries.erase(it);
+    }
+}
+
+void MultiRangeBufferCache::Invalidate(u64 key) {
+    if (auto const it = entries.find(key); it != entries.end()) {
+        it->second.dirty = true;
+    }
+}
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_multi_range_buffer.h

@@ -0,0 +1,105 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <span>
+#include <vector>
+
+#include <boost/container/static_vector.hpp>
+
+#include "common/common_funcs.h"
+#include "common/common_types.h"
+#include "common/container/unordered_map.h"
+#include "video_core/vulkan_common/vulkan_memory_allocator.h"
+#include "video_core/vulkan_common/vulkan_wrapper.h"
+
+namespace Vulkan {
+
+using SparseBuffer = vk::Handle<VkBuffer, VkDevice, vk::DeviceDispatch>;
+
+class Device;
+class Scheduler;
+
+struct MultiRangeSource {
+    VkBuffer handle{};
+    VkDeviceMemory memory{};
+    VkDeviceSize memory_offset{};
+    VkDeviceSize offset{};
+    VkDeviceSize size{};
+    u64 write_tick{};
+    u32 memory_type{};
+};
+
+struct MultiRangeRef {
+    VkBuffer handle{};
+    VkDeviceAddress address{};
+    VkDeviceSize size{};
+    bool sparse{};
+    bool needs_gather{};
+};
+
+class MultiRangeBufferCache final {
+public:
+    static constexpr VkDeviceSize DEFAULT_BLOCK_SIZE = 64 * 1024;
+    static constexpr size_t MAX_RETIRED = 256;
+
+    explicit MultiRangeBufferCache(const Device& device);
+
+    YUZU_NON_COPYABLE(MultiRangeBufferCache);
+
+    [[nodiscard]] MultiRangeRef Get(const Device& device, Scheduler& scheduler,
+                                    MemoryAllocator& memory_allocator, u64 key,
+                                    std::span<const MultiRangeSource> sources,
+                                    VkDeviceSize total);
+
+    void MarkGathered(u64 key);
+
+    void Invalidate(u64 key);
+
+    void DropOwner(Scheduler& scheduler, VkBuffer owner);
+
+    VkDeviceSize block_size{DEFAULT_BLOCK_SIZE};
+    bool use_sparse{};
+
+private:
+    struct Retired {
+        SparseBuffer handle;
+        vk::Buffer gathered;
+        u64 tick{};
+    };
+
+    struct Entry {
+        vk::Buffer gathered;
+        SparseBuffer sparse_handle;
+        std::vector<VkBuffer> owners;
+        VkDeviceAddress address{};
+        VkDeviceSize size{};
+        u64 geometry{};
+        u64 content{};
+        bool dirty{true};
+    };
+
+    [[nodiscard]] u64 HashSources(std::span<const MultiRangeSource> sources) const;
+
+    [[nodiscard]] u64 HashContent(std::span<const MultiRangeSource> sources) const;
+
+    [[nodiscard]] bool CanBindSparse(std::span<const MultiRangeSource> sources) const;
+
+    [[nodiscard]] SparseBuffer CreateSparse(const Device& device, Scheduler& scheduler,
+                                            std::span<const MultiRangeSource> sources,
+                                            VkDeviceSize total);
+
+    [[nodiscard]] VkDeviceSize QueryBlockSize(const Device& device, u32& memory_type_bits) const;
+
+    void RetireEntry(Scheduler& scheduler, Entry& entry);
+
+    void DrainRetired(Scheduler& scheduler);
+
+    ::Common::unordered_map<u64, Entry> entries;
+    boost::container::static_vector<Retired, MAX_RETIRED> retired;
+    u32 sparse_memory_type_bits{};
+    VkBufferUsageFlags sparse_usage{};
+};
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -819,6 +819,7 @@ void RasterizerVulkan::ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) {
         std::scoped_lock lock{texture_cache.mutex};
         texture_cache.UnmapGPUMemory(as_id, addr, size);
     }
+    buffer_cache.UnmapGPUMemory(as_id, addr, size);
 }
 
 void RasterizerVulkan::SignalFence(std::function<void()>&& func) {

src/video_core/vulkan_common/vulkan_device.cpp

@@ -1570,6 +1570,8 @@ void Device::SetupFamilies(VkSurfaceKHR surface) {
     }
     if (graphics) {
         graphics_family = *graphics;
+        graphics_family_sparse_binding =
+            (queue_family_properties[*graphics].queueFlags & VK_QUEUE_SPARSE_BINDING_BIT) != 0;
     }
     if (present) {
         present_family = *present;

src/video_core/vulkan_common/vulkan_device.h

@@ -317,6 +317,10 @@ public:
         return properties.driver.driverID;
     }
 
+    bool IsSparseBindingSupported() const {
+        return features.features.sparseBinding && graphics_family_sparse_binding;
+    }
+
     /// Returns true for tile-based deferred renderers.
     bool IsTiler() const {
         switch (GetDriverID()) {
@@ -1147,6 +1151,7 @@ private:
     u32 instance_version{};      ///< Vulkan instance version.
     u32 graphics_family{};       ///< Main graphics queue family index.
     u32 present_family{};        ///< Main present queue family index.
+    bool graphics_family_sparse_binding{};
 
     struct Extensions {
 #define EXTENSION(prefix, macro_name, var_name) bool var_name{};

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -275,9 +275,63 @@ vk::Buffer MemoryAllocator::CreateBuffer(const VkBufferCreateInfo &ci, MemoryUsa
     const std::span<u8> mapped_data = data ? std::span<u8>{data, ci.size} : std::span<u8>{};
     const bool is_coherent = (property_flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) != 0;
 
-    return vk::Buffer(handle, *device.GetLogical(), allocator, allocation, mapped_data,
-                        is_coherent,
-                        device.GetDispatchLoader());
+    const vk::MemoryLocation location{
+        .memory = alloc_info.deviceMemory,
+        .offset = alloc_info.offset,
+        .memory_type = alloc_info.memoryType,
+    };
+    return vk::Buffer(handle, *device.GetLogical(), allocator, allocation, mapped_data, is_coherent,
+                      location, device.GetDispatchLoader());
+}
+
+vk::Buffer MemoryAllocator::CreateBuffer(const VkBufferCreateInfo &ci, MemoryUsage usage,
+                                         VkDeviceSize min_alignment) const {
+    if (min_alignment <= 1) {
+        return CreateBuffer(ci, usage);
+    }
+    VkMemoryPropertyFlags anv_flags = 0;
+    if (usage == MemoryUsage::Stream &&
+        device.GetDriverID() == VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA) {
+        anv_flags = VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
+    }
+    u32 memory_type_bits = valid_memory_types;
+    if (usage == MemoryUsage::Stream) {
+        memory_type_bits = 0u;
+    }
+    const VmaAllocationCreateInfo alloc_ci = {
+        .flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | MemoryUsageVmaFlags(usage),
+        .usage = MemoryUsageVma(usage),
+        .requiredFlags = 0,
+        .preferredFlags = MemoryUsagePreferredVmaFlags(usage) | anv_flags,
+        .memoryTypeBits = memory_type_bits,
+        .pool = VK_NULL_HANDLE,
+        .pUserData = nullptr,
+        .priority = 0.f,
+    };
+
+    VkBuffer handle{};
+    VmaAllocationInfo alloc_info{};
+    VmaAllocation allocation{};
+    VkMemoryPropertyFlags property_flags{};
+
+    vk::Check(vmaCreateBufferWithAlignment(allocator, &ci, &alloc_ci, min_alignment, &handle,
+                                           &allocation, &alloc_info));
+    vmaGetAllocationMemoryProperties(allocator, allocation, &property_flags);
+
+    u8 *data = reinterpret_cast<u8 *>(alloc_info.pMappedData);
+    std::span<u8> mapped_data{};
+    if (data) {
+        mapped_data = std::span<u8>{data, ci.size};
+    }
+    const bool is_coherent = (property_flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) != 0;
+
+    const vk::MemoryLocation location{
+        .memory = alloc_info.deviceMemory,
+        .offset = alloc_info.offset,
+        .memory_type = alloc_info.memoryType,
+    };
+    return vk::Buffer(handle, *device.GetLogical(), allocator, allocation, mapped_data, is_coherent,
+                      location, device.GetDispatchLoader());
 }
 
 MemoryCommit MemoryAllocator::Commit(const VkMemoryRequirements &reqs, MemoryUsage usage)

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
@@ -107,6 +107,9 @@ namespace Vulkan {
 
         vk::Buffer CreateBuffer(const VkBufferCreateInfo &ci, MemoryUsage usage) const;
 
+        vk::Buffer CreateBuffer(const VkBufferCreateInfo &ci, MemoryUsage usage,
+                                VkDeviceSize min_alignment) const;
+
         /**
          * Commits a memory with the specified requirements.
          *

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -229,6 +229,7 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkGetPipelineExecutableStatisticsKHR);
     X(vkGetSemaphoreCounterValue);
     X(vkMapMemory);
+    X(vkQueueBindSparse);
     X(vkQueueSubmit);
     X(vkQueueSubmit2);
     X(vkResetFences);

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -345,6 +345,7 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkGetQueryPoolResults vkGetQueryPoolResults{};
     PFN_vkGetSemaphoreCounterValue vkGetSemaphoreCounterValue{};
     PFN_vkMapMemory vkMapMemory{};
+    PFN_vkQueueBindSparse vkQueueBindSparse{};
     PFN_vkQueueSubmit vkQueueSubmit{};
     PFN_vkQueueSubmit2 vkQueueSubmit2{};
     PFN_vkResetFences vkResetFences{};
@@ -740,13 +741,20 @@ private:
     const DeviceDispatch* dld = nullptr;
 };
 
+struct MemoryLocation {
+    VkDeviceMemory memory{};
+    VkDeviceSize offset{};
+    u32 memory_type{};
+};
+
 class Buffer {
 public:
     explicit Buffer(VkBuffer handle_, VkDevice owner_, VmaAllocator allocator_,
                     VmaAllocation allocation_, std::span<u8> mapped_, bool is_coherent_,
-                    const DeviceDispatch& dld_) noexcept
+                    MemoryLocation location_, const DeviceDispatch& dld_) noexcept
         : handle{handle_}, owner{owner_}, allocator{allocator_},
-          allocation{allocation_}, mapped{mapped_}, is_coherent{is_coherent_}, dld{&dld_} {}
+          allocation{allocation_}, mapped{mapped_}, location{location_},
+          is_coherent{is_coherent_}, dld{&dld_} {}
     Buffer() = default;
 
     Buffer(const Buffer&) = delete;
@@ -754,7 +762,7 @@ public:
 
     Buffer(Buffer&& rhs) noexcept
         : handle{std::exchange(rhs.handle, VkBuffer{})}, owner{rhs.owner}, allocator{rhs.allocator},
-          allocation{rhs.allocation}, mapped{rhs.mapped},
+          allocation{rhs.allocation}, mapped{rhs.mapped}, location{rhs.location},
           is_coherent{rhs.is_coherent}, dld{rhs.dld} {}
 
     Buffer& operator=(Buffer&& rhs) noexcept {
@@ -764,6 +772,7 @@ public:
         allocator = rhs.allocator;
         allocation = rhs.allocation;
         mapped = rhs.mapped;
+        location = rhs.location;
         is_coherent = rhs.is_coherent;
         dld = rhs.dld;
         return *this;
@@ -811,6 +820,10 @@ public:
 
     void SetObjectNameEXT(const char* name) const;
 
+    MemoryLocation Location() const noexcept {
+        return location;
+    }
+
 private:
     void Release() const noexcept;
 
@@ -819,6 +832,7 @@ private:
     VmaAllocator allocator = nullptr;
     VmaAllocation allocation = nullptr;
     std::span<u8> mapped = {};
+    MemoryLocation location{};
     bool is_coherent = false;
     const DeviceDispatch* dld = nullptr;
 };
@@ -843,6 +857,11 @@ public:
         return dld->vkQueueSubmit2(queue, submit_infos.size(), submit_infos.data(), fence);
     }
 
+    VkResult BindSparse(Span<VkBindSparseInfo> bind_infos,
+                        VkFence fence = VK_NULL_HANDLE) const noexcept {
+        return dld->vkQueueBindSparse(queue, bind_infos.size(), bind_infos.data(), fence);
+    }
+
     VkResult Present(const VkPresentInfoKHR& present_info) const noexcept {
         return dld->vkQueuePresentKHR(queue, &present_info);
     }