Merge pull request #3278 from ReinUsesLisp/vk-memory-manager

renderer_vulkan: Buffer cache, stream buffer and memory manager changes

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -2,124 +2,145 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <algorithm>
 #include <cstring>
 #include <memory>
 #include <optional>
 #include <tuple>
 
-#include "common/alignment.h"
 #include "common/assert.h"
-#include "core/memory.h"
+#include "common/bit_util.h"
-#include "video_core/memory_manager.h"
+#include "core/core.h"
 #include "video_core/renderer_vulkan/declarations.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_device.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_stream_buffer.h"
 
 namespace Vulkan {
 
-CachedBufferEntry::CachedBufferEntry(VAddr cpu_addr, std::size_t size, u64 offset,
+namespace {
-                                     std::size_t alignment, u8* host_ptr)
-    : RasterizerCacheObject{host_ptr}, cpu_addr{cpu_addr}, size{size}, offset{offset},
-      alignment{alignment} {}
 
-VKBufferCache::VKBufferCache(Tegra::MemoryManager& tegra_memory_manager,
+const auto BufferUsage =
-                             Memory::Memory& cpu_memory_,
+    vk::BufferUsageFlagBits::eVertexBuffer | vk::BufferUsageFlagBits::eIndexBuffer |
-                             VideoCore::RasterizerInterface& rasterizer, const VKDevice& device,
+    vk::BufferUsageFlagBits::eUniformBuffer | vk::BufferUsageFlagBits::eStorageBuffer;
-                             VKMemoryManager& memory_manager, VKScheduler& scheduler, u64 size)
+
-    : RasterizerCache{rasterizer}, tegra_memory_manager{tegra_memory_manager}, cpu_memory{
+const auto UploadPipelineStage =
-                                                                                   cpu_memory_} {
+    vk::PipelineStageFlagBits::eTransfer | vk::PipelineStageFlagBits::eVertexInput |
-    const auto usage = vk::BufferUsageFlagBits::eVertexBuffer |
+    vk::PipelineStageFlagBits::eVertexShader | vk::PipelineStageFlagBits::eFragmentShader |
-                       vk::BufferUsageFlagBits::eIndexBuffer |
+    vk::PipelineStageFlagBits::eComputeShader;
-                       vk::BufferUsageFlagBits::eUniformBuffer;
+
-    const auto access = vk::AccessFlagBits::eVertexAttributeRead | vk::AccessFlagBits::eIndexRead |
+const auto UploadAccessBarriers =
-                        vk::AccessFlagBits::eUniformRead;
+    vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eShaderRead |
-    stream_buffer =
+    vk::AccessFlagBits::eUniformRead | vk::AccessFlagBits::eVertexAttributeRead |
-        std::make_unique<VKStreamBuffer>(device, memory_manager, scheduler, size, usage, access,
+    vk::AccessFlagBits::eIndexRead;
-                                         vk::PipelineStageFlagBits::eAllCommands);
+
-    buffer_handle = stream_buffer->GetBuffer();
+auto CreateStreamBuffer(const VKDevice& device, VKScheduler& scheduler) {
+    return std::make_unique<VKStreamBuffer>(device, scheduler, BufferUsage);
 }
 
+} // Anonymous namespace
+
+CachedBufferBlock::CachedBufferBlock(const VKDevice& device, VKMemoryManager& memory_manager,
+                                     CacheAddr cache_addr, std::size_t size)
+    : VideoCommon::BufferBlock{cache_addr, size} {
+    const vk::BufferCreateInfo buffer_ci({}, static_cast<vk::DeviceSize>(size),
+                                         BufferUsage | vk::BufferUsageFlagBits::eTransferSrc |
+                                             vk::BufferUsageFlagBits::eTransferDst,
+                                         vk::SharingMode::eExclusive, 0, nullptr);
+
+    const auto& dld{device.GetDispatchLoader()};
+    const auto dev{device.GetLogical()};
+    buffer.handle = dev.createBufferUnique(buffer_ci, nullptr, dld);
+    buffer.commit = memory_manager.Commit(*buffer.handle, false);
+}
+
+CachedBufferBlock::~CachedBufferBlock() = default;
+
+VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
+                             const VKDevice& device, VKMemoryManager& memory_manager,
+                             VKScheduler& scheduler, VKStagingBufferPool& staging_pool)
+    : VideoCommon::BufferCache<Buffer, vk::Buffer, VKStreamBuffer>{rasterizer, system,
+                                                                   CreateStreamBuffer(device,
+                                                                                      scheduler)},
+      device{device}, memory_manager{memory_manager}, scheduler{scheduler}, staging_pool{
+                                                                                staging_pool} {}
+
 VKBufferCache::~VKBufferCache() = default;
 
-u64 VKBufferCache::UploadMemory(GPUVAddr gpu_addr, std::size_t size, u64 alignment, bool cache) {
+Buffer VKBufferCache::CreateBlock(CacheAddr cache_addr, std::size_t size) {
-    const auto cpu_addr{tegra_memory_manager.GpuToCpuAddress(gpu_addr)};
+    return std::make_shared<CachedBufferBlock>(device, memory_manager, cache_addr, size);
-    ASSERT_MSG(cpu_addr, "Invalid GPU address");
-
-    // Cache management is a big overhead, so only cache entries with a given size.
-    // TODO: Figure out which size is the best for given games.
-    cache &= size >= 2048;
-
-    u8* const host_ptr{cpu_memory.GetPointer(*cpu_addr)};
-    if (cache) {
-        const auto entry = TryGet(host_ptr);
-        if (entry) {
-            if (entry->GetSize() >= size && entry->GetAlignment() == alignment) {
-                return entry->GetOffset();
-            }
-            Unregister(entry);
-        }
-    }
-
-    AlignBuffer(alignment);
-    const u64 uploaded_offset = buffer_offset;
-
-    if (host_ptr == nullptr) {
-        return uploaded_offset;
-    }
-
-    std::memcpy(buffer_ptr, host_ptr, size);
-    buffer_ptr += size;
-    buffer_offset += size;
-
-    if (cache) {
-        auto entry = std::make_shared<CachedBufferEntry>(*cpu_addr, size, uploaded_offset,
-                                                         alignment, host_ptr);
-        Register(entry);
-    }
-
-    return uploaded_offset;
 }
 
-u64 VKBufferCache::UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment) {
+const vk::Buffer* VKBufferCache::ToHandle(const Buffer& buffer) {
-    AlignBuffer(alignment);
+    return buffer->GetHandle();
-    std::memcpy(buffer_ptr, raw_pointer, size);
-    const u64 uploaded_offset = buffer_offset;
-
-    buffer_ptr += size;
-    buffer_offset += size;
-    return uploaded_offset;
 }
 
-std::tuple<u8*, u64> VKBufferCache::ReserveMemory(std::size_t size, u64 alignment) {
+const vk::Buffer* VKBufferCache::GetEmptyBuffer(std::size_t size) {
-    AlignBuffer(alignment);
+    size = std::max(size, std::size_t(4));
-    u8* const uploaded_ptr = buffer_ptr;
+    const auto& empty = staging_pool.GetUnusedBuffer(size, false);
-    const u64 uploaded_offset = buffer_offset;
+    scheduler.RequestOutsideRenderPassOperationContext();
-
+    scheduler.Record([size, buffer = *empty.handle](vk::CommandBuffer cmdbuf, auto& dld) {
-    buffer_ptr += size;
+        cmdbuf.fillBuffer(buffer, 0, size, 0, dld);
-    buffer_offset += size;
+    });
-    return {uploaded_ptr, uploaded_offset};
+    return &*empty.handle;
 }
 
-void VKBufferCache::Reserve(std::size_t max_size) {
+void VKBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-    bool invalidate;
+                                    const u8* data) {
-    std::tie(buffer_ptr, buffer_offset_base, invalidate) = stream_buffer->Reserve(max_size);
+    const auto& staging = staging_pool.GetUnusedBuffer(size, true);
-    buffer_offset = buffer_offset_base;
+    std::memcpy(staging.commit->Map(size), data, size);
 
-    if (invalidate) {
+    scheduler.RequestOutsideRenderPassOperationContext();
-        InvalidateAll();
+    scheduler.Record([staging = *staging.handle, buffer = *buffer->GetHandle(), offset,
-    }
+                      size](auto cmdbuf, auto& dld) {
+        cmdbuf.copyBuffer(staging, buffer, {{0, offset, size}}, dld);
+        cmdbuf.pipelineBarrier(
+            vk::PipelineStageFlagBits::eTransfer, UploadPipelineStage, {}, {},
+            {vk::BufferMemoryBarrier(vk::AccessFlagBits::eTransferWrite, UploadAccessBarriers,
+                                     VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED, buffer,
+                                     offset, size)},
+            {}, dld);
+    });
 }
 
-void VKBufferCache::Send() {
+void VKBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-    stream_buffer->Send(buffer_offset - buffer_offset_base);
+                                      u8* data) {
+    const auto& staging = staging_pool.GetUnusedBuffer(size, true);
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([staging = *staging.handle, buffer = *buffer->GetHandle(), offset,
+                      size](auto cmdbuf, auto& dld) {
+        cmdbuf.pipelineBarrier(
+            vk::PipelineStageFlagBits::eVertexShader | vk::PipelineStageFlagBits::eFragmentShader |
+                vk::PipelineStageFlagBits::eComputeShader,
+            vk::PipelineStageFlagBits::eTransfer, {}, {},
+            {vk::BufferMemoryBarrier(vk::AccessFlagBits::eShaderWrite,
+                                     vk::AccessFlagBits::eTransferRead, VK_QUEUE_FAMILY_IGNORED,
+                                     VK_QUEUE_FAMILY_IGNORED, buffer, offset, size)},
+            {}, dld);
+        cmdbuf.copyBuffer(buffer, staging, {{offset, 0, size}}, dld);
+    });
+    scheduler.Finish();
+
+    std::memcpy(data, staging.commit->Map(size), size);
 }
 
-void VKBufferCache::AlignBuffer(std::size_t alignment) {
+void VKBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
-    // Align the offset, not the mapped pointer
+                              std::size_t dst_offset, std::size_t size) {
-    const u64 offset_aligned = Common::AlignUp(buffer_offset, alignment);
+    scheduler.RequestOutsideRenderPassOperationContext();
-    buffer_ptr += offset_aligned - buffer_offset;
+    scheduler.Record([src_buffer = *src->GetHandle(), dst_buffer = *dst->GetHandle(), src_offset,
-    buffer_offset = offset_aligned;
+                      dst_offset, size](auto cmdbuf, auto& dld) {
+        cmdbuf.copyBuffer(src_buffer, dst_buffer, {{src_offset, dst_offset, size}}, dld);
+        cmdbuf.pipelineBarrier(
+            vk::PipelineStageFlagBits::eTransfer, UploadPipelineStage, {}, {},
+            {vk::BufferMemoryBarrier(vk::AccessFlagBits::eTransferRead,
+                                     vk::AccessFlagBits::eShaderWrite, VK_QUEUE_FAMILY_IGNORED,
+                                     VK_QUEUE_FAMILY_IGNORED, src_buffer, src_offset, size),
+             vk::BufferMemoryBarrier(vk::AccessFlagBits::eTransferWrite, UploadAccessBarriers,
+                                     VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED, dst_buffer,
+                                     dst_offset, size)},
+            {}, dld);
+    });
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -5,105 +5,74 @@
 #pragma once
 
 #include <memory>
-#include <tuple>
+#include <unordered_map>
+#include <vector>
 
 #include "common/common_types.h"
-#include "video_core/gpu.h"
+#include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/rasterizer_cache.h"
 #include "video_core/renderer_vulkan/declarations.h"
-#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_memory_manager.h"
+#include "video_core/renderer_vulkan/vk_resource_manager.h"
+#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
+#include "video_core/renderer_vulkan/vk_stream_buffer.h"
 
-namespace Memory {
+namespace Core {
-class Memory;
+class System;
-}
-
-namespace Tegra {
-class MemoryManager;
 }
 
 namespace Vulkan {
 
 class VKDevice;
-class VKFence;
 class VKMemoryManager;
-class VKStreamBuffer;
+class VKScheduler;
 
-class CachedBufferEntry final : public RasterizerCacheObject {
+class CachedBufferBlock final : public VideoCommon::BufferBlock {
 public:
-    explicit CachedBufferEntry(VAddr cpu_addr, std::size_t size, u64 offset, std::size_t alignment,
+    explicit CachedBufferBlock(const VKDevice& device, VKMemoryManager& memory_manager,
-                               u8* host_ptr);
+                               CacheAddr cache_addr, std::size_t size);
+    ~CachedBufferBlock();
 
-    VAddr GetCpuAddr() const override {
+    const vk::Buffer* GetHandle() const {
-        return cpu_addr;
+        return &*buffer.handle;
-    }
-
-    std::size_t GetSizeInBytes() const override {
-        return size;
-    }
-
-    std::size_t GetSize() const {
-        return size;
-    }
-
-    u64 GetOffset() const {
-        return offset;
-    }
-
-    std::size_t GetAlignment() const {
-        return alignment;
     }
 
 private:
-    VAddr cpu_addr{};
+    VKBuffer buffer;
-    std::size_t size{};
-    u64 offset{};
-    std::size_t alignment{};
 };
 
-class VKBufferCache final : public RasterizerCache<std::shared_ptr<CachedBufferEntry>> {
+using Buffer = std::shared_ptr<CachedBufferBlock>;
+
+class VKBufferCache final : public VideoCommon::BufferCache<Buffer, vk::Buffer, VKStreamBuffer> {
 public:
-    explicit VKBufferCache(Tegra::MemoryManager& tegra_memory_manager, Memory::Memory& cpu_memory_,
+    explicit VKBufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
-                           VideoCore::RasterizerInterface& rasterizer, const VKDevice& device,
+                           const VKDevice& device, VKMemoryManager& memory_manager,
-                           VKMemoryManager& memory_manager, VKScheduler& scheduler, u64 size);
+                           VKScheduler& scheduler, VKStagingBufferPool& staging_pool);
     ~VKBufferCache();
 
-    /// Uploads data from a guest GPU address. Returns host's buffer offset where it's been
+    const vk::Buffer* GetEmptyBuffer(std::size_t size) override;
-    /// allocated.
-    u64 UploadMemory(GPUVAddr gpu_addr, std::size_t size, u64 alignment = 4, bool cache = true);
-
-    /// Uploads from a host memory. Returns host's buffer offset where it's been allocated.
-    u64 UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment = 4);
-
-    /// Reserves memory to be used by host's CPU. Returns mapped address and offset.
-    std::tuple<u8*, u64> ReserveMemory(std::size_t size, u64 alignment = 4);
-
-    /// Reserves a region of memory to be used in subsequent upload/reserve operations.
-    void Reserve(std::size_t max_size);
-
-    /// Ensures that the set data is sent to the device.
-    void Send();
-
-    /// Returns the buffer cache handle.
-    vk::Buffer GetBuffer() const {
-        return buffer_handle;
-    }
 
 protected:
-    // We do not have to flush this cache as things in it are never modified by us.
+    void WriteBarrier() override {}
-    void FlushObjectInner(const std::shared_ptr<CachedBufferEntry>& object) override {}
+
+    Buffer CreateBlock(CacheAddr cache_addr, std::size_t size) override;
+
+    const vk::Buffer* ToHandle(const Buffer& buffer) override;
+
+    void UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                         const u8* data) override;
+
+    void DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                           u8* data) override;
+
+    void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
+                   std::size_t dst_offset, std::size_t size) override;
 
 private:
-    void AlignBuffer(std::size_t alignment);
+    const VKDevice& device;
-
+    VKMemoryManager& memory_manager;
-    Tegra::MemoryManager& tegra_memory_manager;
+    VKScheduler& scheduler;
-    Memory::Memory& cpu_memory;
+    VKStagingBufferPool& staging_pool;
-
-    std::unique_ptr<VKStreamBuffer> stream_buffer;
-    vk::Buffer buffer_handle;
-
-    u8* buffer_ptr = nullptr;
-    u64 buffer_offset = 0;
-    u64 buffer_offset_base = 0;
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_memory_manager.cpp

@@ -6,6 +6,7 @@
 #include <optional>
 #include <tuple>
 #include <vector>
+
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common_types.h"
@@ -16,34 +17,32 @@
 
 namespace Vulkan {
 
-// TODO(Rodrigo): Fine tune this number
+namespace {
-constexpr u64 ALLOC_CHUNK_SIZE = 64 * 1024 * 1024;
+
+u64 GetAllocationChunkSize(u64 required_size) {
+    static constexpr u64 sizes[] = {16ULL << 20, 32ULL << 20, 64ULL << 20, 128ULL << 20};
+    auto it = std::lower_bound(std::begin(sizes), std::end(sizes), required_size);
+    return it != std::end(sizes) ? *it : Common::AlignUp(required_size, 256ULL << 20);
+}
+
+} // Anonymous namespace
 
 class VKMemoryAllocation final {
 public:
     explicit VKMemoryAllocation(const VKDevice& device, vk::DeviceMemory memory,
-                                vk::MemoryPropertyFlags properties, u64 alloc_size, u32 type)
+                                vk::MemoryPropertyFlags properties, u64 allocation_size, u32 type)
-        : device{device}, memory{memory}, properties{properties}, alloc_size{alloc_size},
+        : device{device}, memory{memory}, properties{properties}, allocation_size{allocation_size},
-          shifted_type{ShiftType(type)}, is_mappable{properties &
+          shifted_type{ShiftType(type)} {}
-                                                     vk::MemoryPropertyFlagBits::eHostVisible} {
-        if (is_mappable) {
-            const auto dev = device.GetLogical();
-            const auto& dld = device.GetDispatchLoader();
-            base_address = static_cast<u8*>(dev.mapMemory(memory, 0, alloc_size, {}, dld));
-        }
-    }
 
     ~VKMemoryAllocation() {
         const auto dev = device.GetLogical();
         const auto& dld = device.GetDispatchLoader();
-        if (is_mappable)
-            dev.unmapMemory(memory, dld);
         dev.free(memory, nullptr, dld);
     }
 
     VKMemoryCommit Commit(vk::DeviceSize commit_size, vk::DeviceSize alignment) {
-        auto found = TryFindFreeSection(free_iterator, alloc_size, static_cast<u64>(commit_size),
+        auto found = TryFindFreeSection(free_iterator, allocation_size,
-                                        static_cast<u64>(alignment));
+                                        static_cast<u64>(commit_size), static_cast<u64>(alignment));
         if (!found) {
             found = TryFindFreeSection(0, free_iterator, static_cast<u64>(commit_size),
                                        static_cast<u64>(alignment));
@@ -52,8 +51,7 @@ public:
                 return nullptr;
             }
         }
-        u8* address = is_mappable ? base_address + *found : nullptr;
+        auto commit = std::make_unique<VKMemoryCommitImpl>(device, this, memory, *found,
-        auto commit = std::make_unique<VKMemoryCommitImpl>(this, memory, address, *found,
                                                            *found + commit_size);
         commits.push_back(commit.get());
 
@@ -65,12 +63,10 @@ public:
 
     void Free(const VKMemoryCommitImpl* commit) {
         ASSERT(commit);
-        const auto it =
+
-            std::find_if(commits.begin(), commits.end(),
+        const auto it = std::find(std::begin(commits), std::end(commits), commit);
-                         [&](const auto& stored_commit) { return stored_commit == commit; });
         if (it == commits.end()) {
-            LOG_CRITICAL(Render_Vulkan, "Freeing unallocated commit!");
+            UNREACHABLE_MSG("Freeing unallocated commit!");
-            UNREACHABLE();
             return;
         }
         commits.erase(it);
@@ -88,11 +84,11 @@ private:
     }
 
     /// A memory allocator, it may return a free region between "start" and "end" with the solicited
-    /// requeriments.
+    /// requirements.
     std::optional<u64> TryFindFreeSection(u64 start, u64 end, u64 size, u64 alignment) const {
-        u64 iterator = start;
+        u64 iterator = Common::AlignUp(start, alignment);
-        while (iterator + size < end) {
+        while (iterator + size <= end) {
-            const u64 try_left = Common::AlignUp(iterator, alignment);
+            const u64 try_left = iterator;
             const u64 try_right = try_left + size;
 
             bool overlap = false;
@@ -100,7 +96,7 @@ private:
                 const auto [commit_left, commit_right] = commit->interval;
                 if (try_left < commit_right && commit_left < try_right) {
                     // There's an overlap, continue the search where the overlapping commit ends.
-                    iterator = commit_right;
+                    iterator = Common::AlignUp(commit_right, alignment);
                     overlap = true;
                     break;
                 }
@@ -110,6 +106,7 @@ private:
                 return try_left;
             }
         }
+
         // No free regions where found, return an empty optional.
         return std::nullopt;
     }
@@ -117,12 +114,8 @@ private:
     const VKDevice& device;                   ///< Vulkan device.
     const vk::DeviceMemory memory;            ///< Vulkan memory allocation handler.
     const vk::MemoryPropertyFlags properties; ///< Vulkan properties.
-    const u64 alloc_size;                     ///< Size of this allocation.
+    const u64 allocation_size;                ///< Size of this allocation.
     const u32 shifted_type;                   ///< Stored Vulkan type of this allocation, shifted.
-    const bool is_mappable;                   ///< Whether the allocation is mappable.
-
-    /// Base address of the mapped pointer.
-    u8* base_address{};
 
     /// Hints where the next free region is likely going to be.
     u64 free_iterator{};
@@ -132,13 +125,15 @@ private:
 };
 
 VKMemoryManager::VKMemoryManager(const VKDevice& device)
-    : device{device}, props{device.GetPhysical().getMemoryProperties(device.GetDispatchLoader())},
+    : device{device}, properties{device.GetPhysical().getMemoryProperties(
-      is_memory_unified{GetMemoryUnified(props)} {}
+                          device.GetDispatchLoader())},
+      is_memory_unified{GetMemoryUnified(properties)} {}
 
 VKMemoryManager::~VKMemoryManager() = default;
 
-VKMemoryCommit VKMemoryManager::Commit(const vk::MemoryRequirements& reqs, bool host_visible) {
+VKMemoryCommit VKMemoryManager::Commit(const vk::MemoryRequirements& requirements,
-    ASSERT(reqs.size < ALLOC_CHUNK_SIZE);
+                                       bool host_visible) {
+    const u64 chunk_size = GetAllocationChunkSize(requirements.size);
 
     // When a host visible commit is asked, search for host visible and coherent, otherwise search
     // for a fast device local type.
@@ -147,32 +142,21 @@ VKMemoryCommit VKMemoryManager::Commit(const vk::MemoryRequirements& reqs, bool
             ? vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent
             : vk::MemoryPropertyFlagBits::eDeviceLocal;
 
-    const auto TryCommit = [&]() -> VKMemoryCommit {
+    if (auto commit = TryAllocCommit(requirements, wanted_properties)) {
-        for (auto& alloc : allocs) {
-            if (!alloc->IsCompatible(wanted_properties, reqs.memoryTypeBits))
-                continue;
-
-            if (auto commit = alloc->Commit(reqs.size, reqs.alignment); commit) {
-                return commit;
-            }
-        }
-        return {};
-    };
-
-    if (auto commit = TryCommit(); commit) {
         return commit;
     }
 
     // Commit has failed, allocate more memory.
-    if (!AllocMemory(wanted_properties, reqs.memoryTypeBits, ALLOC_CHUNK_SIZE)) {
+    if (!AllocMemory(wanted_properties, requirements.memoryTypeBits, chunk_size)) {
-        // TODO(Rodrigo): Try to use host memory.
+        // TODO(Rodrigo): Handle these situations in some way like flushing to guest memory.
-        LOG_CRITICAL(Render_Vulkan, "Ran out of memory!");
+        // Allocation has failed, panic.
-        UNREACHABLE();
+        UNREACHABLE_MSG("Ran out of VRAM!");
+        return {};
     }
 
     // Commit again, this time it won't fail since there's a fresh allocation above. If it does,
     // there's a bug.
-    auto commit = TryCommit();
+    auto commit = TryAllocCommit(requirements, wanted_properties);
     ASSERT(commit);
     return commit;
 }
@@ -180,8 +164,7 @@ VKMemoryCommit VKMemoryManager::Commit(const vk::MemoryRequirements& reqs, bool
 VKMemoryCommit VKMemoryManager::Commit(vk::Buffer buffer, bool host_visible) {
     const auto dev = device.GetLogical();
     const auto& dld = device.GetDispatchLoader();
-    const auto requeriments = dev.getBufferMemoryRequirements(buffer, dld);
+    auto commit = Commit(dev.getBufferMemoryRequirements(buffer, dld), host_visible);
-    auto commit = Commit(requeriments, host_visible);
     dev.bindBufferMemory(buffer, commit->GetMemory(), commit->GetOffset(), dld);
     return commit;
 }
@@ -189,25 +172,23 @@ VKMemoryCommit VKMemoryManager::Commit(vk::Buffer buffer, bool host_visible) {
 VKMemoryCommit VKMemoryManager::Commit(vk::Image image, bool host_visible) {
     const auto dev = device.GetLogical();
     const auto& dld = device.GetDispatchLoader();
-    const auto requeriments = dev.getImageMemoryRequirements(image, dld);
+    auto commit = Commit(dev.getImageMemoryRequirements(image, dld), host_visible);
-    auto commit = Commit(requeriments, host_visible);
     dev.bindImageMemory(image, commit->GetMemory(), commit->GetOffset(), dld);
     return commit;
 }
 
 bool VKMemoryManager::AllocMemory(vk::MemoryPropertyFlags wanted_properties, u32 type_mask,
                                   u64 size) {
-    const u32 type = [&]() {
+    const u32 type = [&] {
-        for (u32 type_index = 0; type_index < props.memoryTypeCount; ++type_index) {
+        for (u32 type_index = 0; type_index < properties.memoryTypeCount; ++type_index) {
-            const auto flags = props.memoryTypes[type_index].propertyFlags;
+            const auto flags = properties.memoryTypes[type_index].propertyFlags;
             if ((type_mask & (1U << type_index)) && (flags & wanted_properties)) {
                 // The type matches in type and in the wanted properties.
                 return type_index;
             }
         }
-        LOG_CRITICAL(Render_Vulkan, "Couldn't find a compatible memory type!");
+        UNREACHABLE_MSG("Couldn't find a compatible memory type!");
-        UNREACHABLE();
+        return 0U;
-        return 0u;
     }();
 
     const auto dev = device.GetLogical();
@@ -216,19 +197,33 @@ bool VKMemoryManager::AllocMemory(vk::MemoryPropertyFlags wanted_properties, u32
     // Try to allocate found type.
     const vk::MemoryAllocateInfo memory_ai(size, type);
     vk::DeviceMemory memory;
-    if (const vk::Result res = dev.allocateMemory(&memory_ai, nullptr, &memory, dld);
+    if (const auto res = dev.allocateMemory(&memory_ai, nullptr, &memory, dld);
         res != vk::Result::eSuccess) {
         LOG_CRITICAL(Render_Vulkan, "Device allocation failed with code {}!", vk::to_string(res));
         return false;
     }
-    allocs.push_back(
+    allocations.push_back(
         std::make_unique<VKMemoryAllocation>(device, memory, wanted_properties, size, type));
     return true;
 }
 
-/*static*/ bool VKMemoryManager::GetMemoryUnified(const vk::PhysicalDeviceMemoryProperties& props) {
+VKMemoryCommit VKMemoryManager::TryAllocCommit(const vk::MemoryRequirements& requirements,
-    for (u32 heap_index = 0; heap_index < props.memoryHeapCount; ++heap_index) {
+                                               vk::MemoryPropertyFlags wanted_properties) {
-        if (!(props.memoryHeaps[heap_index].flags & vk::MemoryHeapFlagBits::eDeviceLocal)) {
+    for (auto& allocation : allocations) {
+        if (!allocation->IsCompatible(wanted_properties, requirements.memoryTypeBits)) {
+            continue;
+        }
+        if (auto commit = allocation->Commit(requirements.size, requirements.alignment)) {
+            return commit;
+        }
+    }
+    return {};
+}
+
+/*static*/ bool VKMemoryManager::GetMemoryUnified(
+    const vk::PhysicalDeviceMemoryProperties& properties) {
+    for (u32 heap_index = 0; heap_index < properties.memoryHeapCount; ++heap_index) {
+        if (!(properties.memoryHeaps[heap_index].flags & vk::MemoryHeapFlagBits::eDeviceLocal)) {
             // Memory is considered unified when heaps are device local only.
             return false;
         }
@@ -236,17 +231,28 @@ bool VKMemoryManager::AllocMemory(vk::MemoryPropertyFlags wanted_properties, u32
     return true;
 }
 
-VKMemoryCommitImpl::VKMemoryCommitImpl(VKMemoryAllocation* allocation, vk::DeviceMemory memory,
+VKMemoryCommitImpl::VKMemoryCommitImpl(const VKDevice& device, VKMemoryAllocation* allocation,
-                                       u8* data, u64 begin, u64 end)
+                                       vk::DeviceMemory memory, u64 begin, u64 end)
-    : interval(std::make_pair(begin, end)), memory{memory}, allocation{allocation}, data{data} {}
+    : device{device}, interval{begin, end}, memory{memory}, allocation{allocation} {}
 
 VKMemoryCommitImpl::~VKMemoryCommitImpl() {
     allocation->Free(this);
 }
 
-u8* VKMemoryCommitImpl::GetData() const {
+MemoryMap VKMemoryCommitImpl::Map(u64 size, u64 offset_) const {
-    ASSERT_MSG(data != nullptr, "Trying to access an unmapped commit.");
+    const auto dev = device.GetLogical();
-    return data;
+    const auto address = reinterpret_cast<u8*>(
+        dev.mapMemory(memory, interval.first + offset_, size, {}, device.GetDispatchLoader()));
+    return MemoryMap{this, address};
+}
+
+void VKMemoryCommitImpl::Unmap() const {
+    const auto dev = device.GetLogical();
+    dev.unmapMemory(memory, device.GetDispatchLoader());
+}
+
+MemoryMap VKMemoryCommitImpl::Map() const {
+    return Map(interval.second - interval.first);
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_memory_manager.h

@@ -12,6 +12,7 @@
 
 namespace Vulkan {
 
+class MemoryMap;
 class VKDevice;
 class VKMemoryAllocation;
 class VKMemoryCommitImpl;
@@ -21,11 +22,12 @@ using VKMemoryCommit = std::unique_ptr<VKMemoryCommitImpl>;
 class VKMemoryManager final {
 public:
     explicit VKMemoryManager(const VKDevice& device);
+    VKMemoryManager(const VKMemoryManager&) = delete;
     ~VKMemoryManager();
 
     /**
      * Commits a memory with the specified requeriments.
-     * @param reqs Requeriments returned from a Vulkan call.
+     * @param requirements Requirements returned from a Vulkan call.
      * @param host_visible Signals the allocator that it *must* use host visible and coherent
      *                     memory. When passing false, it will try to allocate device local memory.
      * @returns A memory commit.
@@ -47,25 +49,35 @@ private:
     /// Allocates a chunk of memory.
     bool AllocMemory(vk::MemoryPropertyFlags wanted_properties, u32 type_mask, u64 size);
 
+    /// Tries to allocate a memory commit.
+    VKMemoryCommit TryAllocCommit(const vk::MemoryRequirements& requirements,
+                                  vk::MemoryPropertyFlags wanted_properties);
+
     /// Returns true if the device uses an unified memory model.
-    static bool GetMemoryUnified(const vk::PhysicalDeviceMemoryProperties& props);
+    static bool GetMemoryUnified(const vk::PhysicalDeviceMemoryProperties& properties);
 
     const VKDevice& device;                              ///< Device handler.
-    const vk::PhysicalDeviceMemoryProperties props;          ///< Physical device properties.
+    const vk::PhysicalDeviceMemoryProperties properties; ///< Physical device properties.
     const bool is_memory_unified;                        ///< True if memory model is unified.
-    std::vector<std::unique_ptr<VKMemoryAllocation>> allocs; ///< Current allocations.
+    std::vector<std::unique_ptr<VKMemoryAllocation>> allocations; ///< Current allocations.
 };
 
 class VKMemoryCommitImpl final {
     friend VKMemoryAllocation;
+    friend MemoryMap;
 
 public:
-    explicit VKMemoryCommitImpl(VKMemoryAllocation* allocation, vk::DeviceMemory memory, u8* data,
+    explicit VKMemoryCommitImpl(const VKDevice& device, VKMemoryAllocation* allocation,
-                                u64 begin, u64 end);
+                                vk::DeviceMemory memory, u64 begin, u64 end);
     ~VKMemoryCommitImpl();
 
-    /// Returns the writeable memory map. The commit has to be mappable.
+    /// Maps a memory region and returns a pointer to it.
-    u8* GetData() const;
+    /// It's illegal to have more than one memory map at the same time.
+    MemoryMap Map(u64 size, u64 offset = 0) const;
+
+    /// Maps the whole commit and returns a pointer to it.
+    /// It's illegal to have more than one memory map at the same time.
+    MemoryMap Map() const;
 
     /// Returns the Vulkan memory handler.
     vk::DeviceMemory GetMemory() const {
@@ -78,10 +90,46 @@ public:
     }
 
 private:
+    /// Unmaps memory.
+    void Unmap() const;
+
+    const VKDevice& device;           ///< Vulkan device.
     std::pair<u64, u64> interval{};   ///< Interval where the commit exists.
     vk::DeviceMemory memory;          ///< Vulkan device memory handler.
     VKMemoryAllocation* allocation{}; ///< Pointer to the large memory allocation.
-    u8* data{}; ///< Pointer to the host mapped memory, it has the commit offset included.
+};
+
+/// Holds ownership of a memory map.
+class MemoryMap final {
+public:
+    explicit MemoryMap(const VKMemoryCommitImpl* commit, u8* address)
+        : commit{commit}, address{address} {}
+
+    ~MemoryMap() {
+        if (commit) {
+            commit->Unmap();
+        }
+    }
+
+    /// Prematurely releases the memory map.
+    void Release() {
+        commit->Unmap();
+        commit = nullptr;
+    }
+
+    /// Returns the address of the memory map.
+    u8* GetAddress() const {
+        return address;
+    }
+
+    /// Returns the address of the memory map;
+    operator u8*() const {
+        return address;
+    }
+
+private:
+    const VKMemoryCommitImpl* commit{}; ///< Mapped memory commit.
+    u8* address{};                      ///< Address to the mapped memory.
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_stream_buffer.cpp

@@ -3,86 +3,144 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
-#include <memory>
 #include <optional>
+#include <tuple>
 #include <vector>
 
+#include "common/alignment.h"
 #include "common/assert.h"
 #include "video_core/renderer_vulkan/declarations.h"
 #include "video_core/renderer_vulkan/vk_device.h"
-#include "video_core/renderer_vulkan/vk_memory_manager.h"
 #include "video_core/renderer_vulkan/vk_resource_manager.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_stream_buffer.h"
 
 namespace Vulkan {
 
+namespace {
+
 constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000;
 constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000;
 
-VKStreamBuffer::VKStreamBuffer(const VKDevice& device, VKMemoryManager& memory_manager,
+constexpr u64 STREAM_BUFFER_SIZE = 256 * 1024 * 1024;
-                               VKScheduler& scheduler, u64 size, vk::BufferUsageFlags usage,
+
-                               vk::AccessFlags access, vk::PipelineStageFlags pipeline_stage)
+std::optional<u32> FindMemoryType(const VKDevice& device, u32 filter,
-    : device{device}, scheduler{scheduler}, buffer_size{size}, access{access}, pipeline_stage{
+                                  vk::MemoryPropertyFlags wanted) {
-                                                                                   pipeline_stage} {
+    const auto properties = device.GetPhysical().getMemoryProperties(device.GetDispatchLoader());
-    CreateBuffers(memory_manager, usage);
+    for (u32 i = 0; i < properties.memoryTypeCount; i++) {
-    ReserveWatches(WATCHES_INITIAL_RESERVE);
+        if (!(filter & (1 << i))) {
+            continue;
+        }
+        if ((properties.memoryTypes[i].propertyFlags & wanted) == wanted) {
+            return i;
+        }
+    }
+    return {};
+}
+
+} // Anonymous namespace
+
+VKStreamBuffer::VKStreamBuffer(const VKDevice& device, VKScheduler& scheduler,
+                               vk::BufferUsageFlags usage)
+    : device{device}, scheduler{scheduler} {
+    CreateBuffers(usage);
+    ReserveWatches(current_watches, WATCHES_INITIAL_RESERVE);
+    ReserveWatches(previous_watches, WATCHES_INITIAL_RESERVE);
 }
 
 VKStreamBuffer::~VKStreamBuffer() = default;
 
-std::tuple<u8*, u64, bool> VKStreamBuffer::Reserve(u64 size) {
+std::tuple<u8*, u64, bool> VKStreamBuffer::Map(u64 size, u64 alignment) {
-    ASSERT(size <= buffer_size);
+    ASSERT(size <= STREAM_BUFFER_SIZE);
     mapped_size = size;
 
-    if (offset + size > buffer_size) {
+    if (alignment > 0) {
-        // The buffer would overflow, save the amount of used buffers, signal an invalidation and
+        offset = Common::AlignUp(offset, alignment);
-        // reset the state.
+    }
-        invalidation_mark = used_watches;
+
-        used_watches = 0;
+    WaitPendingOperations(offset);
+
+    bool invalidated = false;
+    if (offset + size > STREAM_BUFFER_SIZE) {
+        // The buffer would overflow, save the amount of used watches and reset the state.
+        invalidation_mark = current_watch_cursor;
+        current_watch_cursor = 0;
         offset = 0;
-    }
 
-    return {mapped_pointer + offset, offset, invalidation_mark.has_value()};
+        // Swap watches and reset waiting cursors.
-}
+        std::swap(previous_watches, current_watches);
+        wait_cursor = 0;
+        wait_bound = 0;
 
-void VKStreamBuffer::Send(u64 size) {
+        // Ensure that we don't wait for uncommitted fences.
-    ASSERT_MSG(size <= mapped_size, "Reserved size is too small");
-
-    if (invalidation_mark) {
-        // TODO(Rodrigo): Find a better way to invalidate than waiting for all watches to finish.
         scheduler.Flush();
-        std::for_each(watches.begin(), watches.begin() + *invalidation_mark,
+
-                      [&](auto& resource) { resource->Wait(); });
+        invalidated = true;
-        invalidation_mark = std::nullopt;
     }
 
-    if (used_watches + 1 >= watches.size()) {
-        // Ensure that there are enough watches.
-        ReserveWatches(WATCHES_RESERVE_CHUNK);
-    }
-    // Add a watch for this allocation.
-    watches[used_watches++]->Watch(scheduler.GetFence());
-
-    offset += size;
-}
-
-void VKStreamBuffer::CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage) {
-    const vk::BufferCreateInfo buffer_ci({}, buffer_size, usage, vk::SharingMode::eExclusive, 0,
-                                         nullptr);
-
     const auto dev = device.GetLogical();
     const auto& dld = device.GetDispatchLoader();
-    buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
+    const auto pointer = reinterpret_cast<u8*>(dev.mapMemory(*memory, offset, size, {}, dld));
-    commit = memory_manager.Commit(*buffer, true);
+    return {pointer, offset, invalidated};
-    mapped_pointer = commit->GetData();
 }
 
-void VKStreamBuffer::ReserveWatches(std::size_t grow_size) {
+void VKStreamBuffer::Unmap(u64 size) {
-    const std::size_t previous_size = watches.size();
+    ASSERT_MSG(size <= mapped_size, "Reserved size is too small");
-    watches.resize(previous_size + grow_size);
+
-    std::generate(watches.begin() + previous_size, watches.end(),
+    const auto dev = device.GetLogical();
-                  []() { return std::make_unique<VKFenceWatch>(); });
+    dev.unmapMemory(*memory, device.GetDispatchLoader());
+
+    offset += size;
+
+    if (current_watch_cursor + 1 >= current_watches.size()) {
+        // Ensure that there are enough watches.
+        ReserveWatches(current_watches, WATCHES_RESERVE_CHUNK);
+    }
+    auto& watch = current_watches[current_watch_cursor++];
+    watch.upper_bound = offset;
+    watch.fence.Watch(scheduler.GetFence());
+}
+
+void VKStreamBuffer::CreateBuffers(vk::BufferUsageFlags usage) {
+    const vk::BufferCreateInfo buffer_ci({}, STREAM_BUFFER_SIZE, usage, vk::SharingMode::eExclusive,
+                                         0, nullptr);
+    const auto dev = device.GetLogical();
+    const auto& dld = device.GetDispatchLoader();
+    buffer = dev.createBufferUnique(buffer_ci, nullptr, dld);
+
+    const auto requirements = dev.getBufferMemoryRequirements(*buffer, dld);
+    // Prefer device local host visible allocations (this should hit AMD's pinned memory).
+    auto type = FindMemoryType(device, requirements.memoryTypeBits,
+                               vk::MemoryPropertyFlagBits::eHostVisible |
+                                   vk::MemoryPropertyFlagBits::eHostCoherent |
+                                   vk::MemoryPropertyFlagBits::eDeviceLocal);
+    if (!type) {
+        // Otherwise search for a host visible allocation.
+        type = FindMemoryType(device, requirements.memoryTypeBits,
+                              vk::MemoryPropertyFlagBits::eHostVisible |
+                                  vk::MemoryPropertyFlagBits::eHostCoherent);
+        ASSERT_MSG(type, "No host visible and coherent memory type found");
+    }
+    const vk::MemoryAllocateInfo alloc_ci(requirements.size, *type);
+    memory = dev.allocateMemoryUnique(alloc_ci, nullptr, dld);
+
+    dev.bindBufferMemory(*buffer, *memory, 0, dld);
+}
+
+void VKStreamBuffer::ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size) {
+    watches.resize(watches.size() + grow_size);
+}
+
+void VKStreamBuffer::WaitPendingOperations(u64 requested_upper_bound) {
+    if (!invalidation_mark) {
+        return;
+    }
+    while (requested_upper_bound < wait_bound && wait_cursor < *invalidation_mark) {
+        auto& watch = previous_watches[wait_cursor];
+        wait_bound = watch.upper_bound;
+        watch.fence.Wait();
+        ++wait_cursor;
+    }
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_stream_buffer.h

@@ -4,28 +4,24 @@
 
 #pragma once
 
-#include <memory>
 #include <optional>
 #include <tuple>
 #include <vector>
 
 #include "common/common_types.h"
 #include "video_core/renderer_vulkan/declarations.h"
-#include "video_core/renderer_vulkan/vk_memory_manager.h"
 
 namespace Vulkan {
 
 class VKDevice;
 class VKFence;
 class VKFenceWatch;
-class VKResourceManager;
 class VKScheduler;
 
-class VKStreamBuffer {
+class VKStreamBuffer final {
 public:
-    explicit VKStreamBuffer(const VKDevice& device, VKMemoryManager& memory_manager,
+    explicit VKStreamBuffer(const VKDevice& device, VKScheduler& scheduler,
-                            VKScheduler& scheduler, u64 size, vk::BufferUsageFlags usage,
+                            vk::BufferUsageFlags usage);
-                            vk::AccessFlags access, vk::PipelineStageFlags pipeline_stage);
     ~VKStreamBuffer();
 
     /**
@@ -34,39 +30,47 @@ public:
      * @returns A tuple in the following order: Raw memory pointer (with offset added), buffer
      * offset and a boolean that's true when buffer has been invalidated.
      */
-    std::tuple<u8*, u64, bool> Reserve(u64 size);
+    std::tuple<u8*, u64, bool> Map(u64 size, u64 alignment);
 
     /// Ensures that "size" bytes of memory are available to the GPU, potentially recording a copy.
-    void Send(u64 size);
+    void Unmap(u64 size);
 
-    vk::Buffer GetBuffer() const {
+    vk::Buffer GetHandle() const {
         return *buffer;
     }
 
 private:
+    struct Watch final {
+        VKFenceWatch fence;
+        u64 upper_bound{};
+    };
+
     /// Creates Vulkan buffer handles committing the required the required memory.
-    void CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage);
+    void CreateBuffers(vk::BufferUsageFlags usage);
 
     /// Increases the amount of watches available.
-    void ReserveWatches(std::size_t grow_size);
+    void ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size);
+
+    void WaitPendingOperations(u64 requested_upper_bound);
 
     const VKDevice& device;                      ///< Vulkan device manager.
     VKScheduler& scheduler;                      ///< Command scheduler.
-    const u64 buffer_size;                       ///< Total size of the stream buffer.
     const vk::AccessFlags access;                ///< Access usage of this stream buffer.
     const vk::PipelineStageFlags pipeline_stage; ///< Pipeline usage of this stream buffer.
 
     UniqueBuffer buffer;       ///< Mapped buffer.
-    VKMemoryCommit commit; ///< Memory commit.
+    UniqueDeviceMemory memory; ///< Memory allocation.
-    u8* mapped_pointer{};  ///< Pointer to the host visible commit
 
     u64 offset{};      ///< Buffer iterator.
     u64 mapped_size{}; ///< Size reserved for the current copy.
 
-    std::vector<std::unique_ptr<VKFenceWatch>> watches; ///< Total watches
+    std::vector<Watch> current_watches;           ///< Watches recorded in the current iteration.
-    std::size_t used_watches{}; ///< Count of watches, reset on invalidation.
+    std::size_t current_watch_cursor{};           ///< Count of watches, reset on invalidation.
-    std::optional<std::size_t>
+    std::optional<std::size_t> invalidation_mark; ///< Number of watches used in the previous cycle.
-        invalidation_mark{}; ///< Number of watches used in the current invalidation.
+
+    std::vector<Watch> previous_watches; ///< Watches used in the previous iteration.
+    std::size_t wait_cursor{};           ///< Last watch being waited for completion.
+    u64 wait_bound{};                    ///< Highest offset being watched for completion.
 };
 
 } // namespace Vulkan