New resolution on the multirange resolve

src/video_core/buffer_cache/buffer_cache.h

@@ -1004,37 +1004,66 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
 }
 
 template <class P>
-bool BufferCache<P>::BindMultiRangeStorage(const Binding& binding, bool is_written) {
+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{}); }) {
         if (binding.gpu_addr == 0 || binding.size == 0) {
-            return false;
+            return;
         }
         if (is_written && !runtime.PrefersSparseSources()) {
-            return false;
+            return;
         }
-        const VirtualSegments* segments =
+        const VirtualSegments* found =
             virtual_ranges.Query(*gpu_memory, binding.gpu_addr, binding.size);
-        if (!segments || segments->size() < 2) {
-            return false;
+        if (!found || found->size() < 2) {
+            return;
         }
-        const u64 key = (static_cast<u64>(gpu_memory->GetID()) << 48) ^ binding.gpu_addr;
+        const VirtualSegments segments = *found;
+        const u32 first = static_cast<u32>(pool.size());
         const bool prefer_sparse = runtime.PrefersSparseSources();
-        runtime.ResetMultiRange();
-        for (const VirtualSegment& segment : *segments) {
+        for (const VirtualSegment& segment : segments) {
             const BufferId buffer_id =
                 FindBuffer(segment.device_addr, segment.size, prefer_sparse);
             if (!buffer_id) {
-                return false;
+                pool.resize(first);
+                return;
             }
-            Buffer& buffer = slot_buffers[buffer_id];
-            TouchBuffer(buffer, buffer_id);
+            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{}); }) {
+        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(buffer_id, segment.device_addr, segment.size);
+                MarkWrittenBuffer(segment.buffer_id, segment.device_addr, segment.size);
             }
             runtime.PushMultiRangeSource(buffer, offset, segment.size);
         }
@@ -1050,7 +1079,7 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
     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)) {
+        if (BindMultiRangeStorage(binding, is_written, graphics_segments)) {
             return;
         }
         Buffer& buffer = slot_buffers[binding.buffer_id];
@@ -1190,7 +1219,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
         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)) {
+        if (BindMultiRangeStorage(binding, is_written, compute_segments)) {
             return;
         }
         Buffer& buffer = slot_buffers[binding.buffer_id];
@@ -1245,6 +1274,7 @@ void BufferCache<P>::BindHostComputeTextureBuffers() {
 
 template <class P>
 void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) {
+    graphics_segments.clear();
     BufferOperations([&]() {
         if (is_indexed) {
             UpdateIndexBuffer();
@@ -1264,6 +1294,7 @@ void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) {
 
 template <class P>
 void BufferCache<P>::DoUpdateComputeBuffers() {
+    compute_segments.clear();
     BufferOperations([&]() {
         UpdateComputeUniformBuffers();
         UpdateComputeStorageBuffers();
@@ -1406,6 +1437,8 @@ void BufferCache<P>::UpdateStorageBuffers(size_t stage) {
         Binding& binding = channel_state->storage_buffers[stage][index];
         const BufferId buffer_id = FindBuffer(binding.device_addr, binding.size);
         binding.buffer_id = buffer_id;
+        const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
+        ResolveMultiRangeStorage(binding, is_written, graphics_segments);
     });
 }
 
@@ -1469,6 +1502,9 @@ void BufferCache<P>::UpdateComputeStorageBuffers() {
         // Resolve buffer
         Binding& binding = channel_state->compute_storage_buffers[index];
         binding.buffer_id = FindBuffer(binding.device_addr, binding.size);
+        const bool is_written =
+            ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
+        ResolveMultiRangeStorage(binding, is_written, compute_segments);
     });
 }
 

src/video_core/buffer_cache/buffer_cache_base.h

@@ -85,6 +85,14 @@ struct Binding {
     u32 size{};
     BufferId buffer_id;
     GPUVAddr gpu_addr{};
+    u32 segment_first{};
+    u32 segment_count{};
+};
+
+struct MultiRangeSegment {
+    BufferId buffer_id;
+    DAddr device_addr{};
+    u32 size{};
 };
 
 struct TextureBufferBinding : Binding {
@@ -217,7 +225,11 @@ public:
 
     void TickFrame();
 
-    bool BindMultiRangeStorage(const Binding& binding, bool is_written);
+    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);
 
@@ -523,6 +535,8 @@ private:
     Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
     u64 frame_tick = 0;
     VirtualRangeCache virtual_ranges;
+    std::vector<MultiRangeSegment> graphics_segments;
+    std::vector<MultiRangeSegment> compute_segments;
     u64 total_used_memory = 0;
     u64 minimum_memory = 0;
     u64 critical_memory = 0;

src/video_core/buffer_cache/virtual_range_cache.h

@@ -43,20 +43,31 @@ public:
         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) {
-            const std::optional<DAddr> device_addr = memory.GpuToCpuAddress(range_addr);
-            if (!device_addr) {
+            if (range_addr != expected || range_size == 0) {
+                contiguous = false;
                 break;
             }
-            u32 segment_size = (std::numeric_limits<u32>::max)();
-            if (range_size < static_cast<size_t>(segment_size)) {
-                segment_size = static_cast<u32>(range_size);
+            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 = segment_size,
+                .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;

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -192,6 +192,7 @@ public:
             .memory_offset = location.offset,
             .offset = offset,
             .size = size,
+            .memory_type = location.memory_type,
         });
         multi_range_total += size;
     }

src/video_core/renderer_vulkan/vk_multi_range_buffer.cpp

@@ -21,11 +21,13 @@ MultiRangeBufferCache::MultiRangeBufferCache(const Device& device_,
     if (!device.IsSparseBindingSupported()) {
         return;
     }
-    const VkDeviceSize queried = QueryBlockSize();
-    if (queried == 0) {
+    u32 memory_type_bits = 0;
+    const VkDeviceSize queried = QueryBlockSize(memory_type_bits);
+    if (queried == 0 || memory_type_bits == 0) {
         return;
     }
     block_size = queried;
+    sparse_memory_type_bits = memory_type_bits;
     use_sparse = true;
 }
 
@@ -44,7 +46,7 @@ MultiRangeBufferCache::~MultiRangeBufferCache() {
     retired.clear();
 }
 
-VkDeviceSize MultiRangeBufferCache::QueryBlockSize() const {
+VkDeviceSize MultiRangeBufferCache::QueryBlockSize(u32& memory_type_bits) const {
     const VkDevice logical = *device.GetLogical();
     const auto& dld = device.GetDispatchLoader();
     const VkBufferCreateInfo probe_ci{
@@ -73,6 +75,7 @@ VkDeviceSize MultiRangeBufferCache::QueryBlockSize() const {
     };
     dld.vkGetBufferMemoryRequirements2(logical, &reqs_info, &reqs2);
     dld.vkDestroyBuffer(logical, probe, nullptr);
+    memory_type_bits = reqs2.memoryRequirements.memoryTypeBits;
     return reqs2.memoryRequirements.alignment;
 }
 
@@ -98,6 +101,12 @@ bool MultiRangeBufferCache::CanBindSparse(std::span<const MultiRangeSource> sour
         if (source.memory == VK_NULL_HANDLE) {
             return false;
         }
+        if (source.memory_type >= 32) {
+            return false;
+        }
+        if (((sparse_memory_type_bits >> source.memory_type) & 1) == 0) {
+            return false;
+        }
         const VkDeviceSize memory_offset = source.memory_offset + source.offset;
         if ((memory_offset % block_size) != 0) {
             return false;

src/video_core/renderer_vulkan/vk_multi_range_buffer.h

@@ -22,6 +22,7 @@ struct MultiRangeSource {
     VkDeviceSize memory_offset{};
     VkDeviceSize offset{};
     VkDeviceSize size{};
+    u32 memory_type{};
 };
 
 struct MultiRangeRef {
@@ -81,7 +82,7 @@ private:
     [[nodiscard]] VkBuffer CreateSparse(std::span<const MultiRangeSource> sources,
                                         VkDeviceSize total);
 
-    [[nodiscard]] VkDeviceSize QueryBlockSize() const;
+    [[nodiscard]] VkDeviceSize QueryBlockSize(u32& memory_type_bits) const;
 
     void DestroySparse(VkBuffer handle);
 
@@ -92,6 +93,7 @@ private:
     Scheduler& scheduler;
     bool use_sparse{};
     VkDeviceSize block_size{DEFAULT_BLOCK_SIZE};
+    u32 sparse_memory_type_bits{};
     VkBufferUsageFlags sparse_usage{};
     std::unordered_map<u64, Entry> entries;
     std::vector<Retired> retired;

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -549,6 +549,7 @@ MemoryLocation Buffer::Location() const noexcept {
     return MemoryLocation{
         .memory = info.deviceMemory,
         .offset = info.offset,
+        .memory_type = info.memoryType,
     };
 }
 

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -744,6 +744,7 @@ private:
 struct MemoryLocation {
     VkDeviceMemory memory{};
     VkDeviceSize offset{};
+    u32 memory_type{};
 };
 
 class Buffer {