codex-5.2: ресурсы

Src/Client/Vulkan/AtlasPipeline/PipelinedTextureAtlas.cpp

@@ -0,0 +1,198 @@
+#include "PipelinedTextureAtlas.hpp"
+
+PipelinedTextureAtlas::PipelinedTextureAtlas(TextureAtlas&& tk)
+    : Super(std::move(tk)) {}
+
+PipelinedTextureAtlas::AtlasTextureId PipelinedTextureAtlas::getByPipeline(const HashedPipeline& pipeline) {
+    auto iter = _PipeToTexId.find(pipeline);
+    if (iter == _PipeToTexId.end()) {
+        AtlasTextureId atlasTexId = Super.registerTexture();
+        _PipeToTexId.insert({pipeline, atlasTexId});
+        _ChangedPipelines.push_back(pipeline);
+
+        for (uint32_t texId : pipeline.getDependencedTextures()) {
+            _AddictedTextures[texId].push_back(pipeline);
+        }
+
+        return atlasTexId;
+    }
+
+    return iter->second;
+}
+
+void PipelinedTextureAtlas::freeByPipeline(const HashedPipeline& pipeline) {
+    auto iter = _PipeToTexId.find(pipeline);
+    if (iter == _PipeToTexId.end()) {
+        return;
+    }
+
+    for (uint32_t texId : pipeline.getDependencedTextures()) {
+        auto iterAT = _AddictedTextures.find(texId);
+        assert(iterAT != _AddictedTextures.end());
+        auto iterATSub = std::find(iterAT->second.begin(), iterAT->second.end(), pipeline);
+        assert(iterATSub != iterAT->second.end());
+        iterAT->second.erase(iterATSub);
+    }
+
+    Super.removeTexture(iter->second);
+    _AtlasCpuTextures.erase(iter->second);
+    _PipeToTexId.erase(iter);
+}
+
+void PipelinedTextureAtlas::updateTexture(uint32_t texId, const StoredTexture& texture) {
+    _ResToTexture[texId] = texture;
+    _ChangedTextures.push_back(texId);
+}
+
+void PipelinedTextureAtlas::updateTexture(uint32_t texId, StoredTexture&& texture) {
+    _ResToTexture[texId] = std::move(texture);
+    _ChangedTextures.push_back(texId);
+}
+
+void PipelinedTextureAtlas::freeTexture(uint32_t texId) {
+    auto iter = _ResToTexture.find(texId);
+    if (iter != _ResToTexture.end()) {
+        _ResToTexture.erase(iter);
+    }
+}
+
+bool PipelinedTextureAtlas::getHostTexture(TextureId texId, HostTextureView& out) const {
+    auto fill = [&](const StoredTexture& tex) -> bool {
+        if (tex._Pixels.empty() || tex._Widht == 0 || tex._Height == 0) {
+            return false;
+        }
+        out.width = tex._Widht;
+        out.height = tex._Height;
+        out.rowPitchBytes = static_cast<uint32_t>(tex._Widht) * 4u;
+        out.pixelsRGBA8 = reinterpret_cast<const uint8_t*>(tex._Pixels.data());
+        return true;
+    };
+
+    auto it = _ResToTexture.find(texId);
+    if (it != _ResToTexture.end() && fill(it->second)) {
+        return true;
+    }
+
+    auto itAtlas = _AtlasCpuTextures.find(texId);
+    if (itAtlas != _AtlasCpuTextures.end() && fill(itAtlas->second)) {
+        return true;
+    }
+    return false;
+}
+
+StoredTexture PipelinedTextureAtlas::_generatePipelineTexture(const HashedPipeline& pipeline) {
+    std::vector<detail::Word> words(pipeline._Pipeline.begin(), pipeline._Pipeline.end());
+    if (words.empty()) {
+        if (auto tex = tryCopyFirstDependencyTexture(pipeline)) {
+            return *tex;
+        }
+        return makeSolidColorTexture(0xFFFF00FFu);
+    }
+
+    TexturePipelineProgram program;
+    program.fromWords(std::move(words));
+
+    TexturePipelineProgram::OwnedTexture baked;
+    auto provider = [this](uint32_t texId) -> std::optional<Texture> {
+        auto iter = _ResToTexture.find(texId);
+        if (iter == _ResToTexture.end()) {
+            return std::nullopt;
+        }
+        const StoredTexture& stored = iter->second;
+        if (stored._Pixels.empty() || stored._Widht == 0 || stored._Height == 0) {
+            return std::nullopt;
+        }
+        Texture tex{};
+        tex.Width = stored._Widht;
+        tex.Height = stored._Height;
+        tex.Pixels = stored._Pixels.data();
+        return tex;
+    };
+
+    if (!program.bake(provider, baked, nullptr)) {
+        if (auto tex = tryCopyFirstDependencyTexture(pipeline)) {
+            return *tex;
+        }
+        return makeSolidColorTexture(0xFFFF00FFu);
+    }
+
+    const uint32_t width = baked.Width;
+    const uint32_t height = baked.Height;
+    if (width == 0 || height == 0 ||
+        width > std::numeric_limits<uint16_t>::max() ||
+        height > std::numeric_limits<uint16_t>::max() ||
+        baked.Pixels.size() != static_cast<size_t>(width) * static_cast<size_t>(height)) {
+        if (auto tex = tryCopyFirstDependencyTexture(pipeline)) {
+            return *tex;
+        }
+        return makeSolidColorTexture(0xFFFF00FFu);
+    }
+
+    return StoredTexture(static_cast<uint16_t>(width),
+                         static_cast<uint16_t>(height),
+                         std::move(baked.Pixels));
+}
+
+void PipelinedTextureAtlas::flushNewPipelines() {
+    std::vector<uint32_t> changedTextures = std::move(_ChangedTextures);
+
+    std::sort(changedTextures.begin(), changedTextures.end());
+    changedTextures.erase(std::unique(changedTextures.begin(), changedTextures.end()), changedTextures.end());
+
+    std::vector<HashedPipeline> changedPipelineTextures;
+    for (uint32_t texId : changedTextures) {
+        auto iter = _AddictedTextures.find(texId);
+        if (iter == _AddictedTextures.end()) {
+            continue;
+        }
+
+        changedPipelineTextures.append_range(iter->second);
+    }
+
+    changedPipelineTextures.append_range(std::move(_ChangedPipelines));
+    changedTextures.clear();
+
+    std::sort(changedPipelineTextures.begin(), changedPipelineTextures.end());
+    changedPipelineTextures.erase(std::unique(changedPipelineTextures.begin(), changedPipelineTextures.end()),
+                                  changedPipelineTextures.end());
+
+    for (const HashedPipeline& pipeline : changedPipelineTextures) {
+        auto iterPTTI = _PipeToTexId.find(pipeline);
+        assert(iterPTTI != _PipeToTexId.end());
+
+        StoredTexture texture = _generatePipelineTexture(pipeline);
+        AtlasTextureId atlasTexId = iterPTTI->second;
+        auto& stored = _AtlasCpuTextures[atlasTexId];
+        stored = std::move(texture);
+        if (!stored._Pixels.empty()) {
+            Super.setTextureData(atlasTexId,
+                                 stored._Widht,
+                                 stored._Height,
+                                 stored._Pixels.data(),
+                                 stored._Widht * 4u);
+        }
+    }
+}
+
+TextureAtlas::DescriptorOut PipelinedTextureAtlas::flushUploadsAndBarriers(VkCommandBuffer cmdBuffer) {
+    return Super.flushUploadsAndBarriers(cmdBuffer);
+}
+
+void PipelinedTextureAtlas::notifyGpuFinished() {
+    Super.notifyGpuFinished();
+}
+
+std::optional<StoredTexture> PipelinedTextureAtlas::tryCopyFirstDependencyTexture(const HashedPipeline& pipeline) const {
+    auto deps = pipeline.getDependencedTextures();
+    if (!deps.empty()) {
+        auto iter = _ResToTexture.find(deps.front());
+        if (iter != _ResToTexture.end()) {
+            return iter->second;
+        }
+    }
+    return std::nullopt;
+}
+
+StoredTexture PipelinedTextureAtlas::makeSolidColorTexture(uint32_t rgba) {
+    return StoredTexture(1, 1, std::vector<uint32_t>{rgba});
+}

Src/Client/Vulkan/AtlasPipeline/PipelinedTextureAtlas.hpp

@@ -0,0 +1,380 @@
+#pragma once
+
+#include "TextureAtlas.hpp"
+#include "TexturePipelineProgram.hpp"
+#include <algorithm>
+#include <cassert>
+#include <cstdlib>
+#include <cstring>
+#include <optional>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "boost/container/small_vector.hpp"
+
+using TextureId = uint32_t;
+
+namespace detail {
+
+using Word = TexturePipelineProgram::Word;
+
+enum class Op16 : Word {
+    End        = 0,
+    Base_Tex   = 1,
+    Base_Fill  = 2,
+    Resize     = 10,
+    Transform  = 11,
+    Opacity    = 12,
+    NoAlpha    = 13,
+    MakeAlpha  = 14,
+    Invert     = 15,
+    Brighten   = 16,
+    Contrast   = 17,
+    Multiply   = 18,
+    Screen     = 19,
+    Colorize   = 20,
+    Overlay    = 30,
+    Mask       = 31,
+    LowPart    = 32,
+    Combine    = 40
+};
+
+enum class SrcKind16 : Word { TexId = 0, Sub = 1 };
+
+struct SrcRef16 {
+    SrcKind16 kind{SrcKind16::TexId};
+    Word a = 0;
+    Word b = 0;
+};
+
+inline uint32_t makeU32(Word lo, Word hi) {
+    return uint32_t(lo) | (uint32_t(hi) << 16);
+}
+
+inline void addUniqueDep(boost::container::small_vector<uint32_t, 8>& deps, uint32_t id) {
+    if (id == TextureAtlas::kOverflowId) {
+        return;
+    }
+    if (std::find(deps.begin(), deps.end(), id) == deps.end()) {
+        deps.push_back(id);
+    }
+}
+
+inline bool readSrc(const std::vector<Word>& words, size_t end, size_t& ip, SrcRef16& out) {
+    if (ip + 2 >= end) {
+        return false;
+    }
+    out.kind = static_cast<SrcKind16>(words[ip++]);
+    out.a = words[ip++];
+    out.b = words[ip++];
+    return true;
+}
+
+inline void extractPipelineDependencies(const std::vector<Word>& words,
+                                        size_t start,
+                                        size_t end,
+                                        boost::container::small_vector<uint32_t, 8>& deps,
+                                        std::vector<std::pair<size_t, size_t>>& visited) {
+    if (start >= end || end > words.size()) {
+        return;
+    }
+    const std::pair<size_t, size_t> key{start, end};
+    if (std::find(visited.begin(), visited.end(), key) != visited.end()) {
+        return;
+    }
+    visited.push_back(key);
+
+    size_t ip = start;
+    auto need = [&](size_t n) { return ip + n <= end; };
+    auto handleSrc = [&](const SrcRef16& src) {
+        if (src.kind == SrcKind16::TexId) {
+            addUniqueDep(deps, makeU32(src.a, src.b));
+            return;
+        }
+        if (src.kind == SrcKind16::Sub) {
+            size_t subStart = static_cast<size_t>(src.a);
+            size_t subEnd = subStart + static_cast<size_t>(src.b);
+            if (subStart < subEnd && subEnd <= words.size()) {
+                extractPipelineDependencies(words, subStart, subEnd, deps, visited);
+            }
+        }
+    };
+
+    while (ip < end) {
+        if (!need(1)) break;
+        Op16 op = static_cast<Op16>(words[ip++]);
+        switch (op) {
+            case Op16::End:
+                return;
+
+            case Op16::Base_Tex: {
+                if (!need(3)) return;
+                SrcRef16 src{};
+                if (!readSrc(words, end, ip, src)) return;
+                handleSrc(src);
+            } break;
+
+            case Op16::Base_Fill:
+                if (!need(4)) return;
+                ip += 4;
+                break;
+
+            case Op16::Overlay:
+            case Op16::Mask: {
+                if (!need(3)) return;
+                SrcRef16 src{};
+                if (!readSrc(words, end, ip, src)) return;
+                handleSrc(src);
+            } break;
+
+            case Op16::LowPart: {
+                if (!need(1 + 3)) return;
+                ip += 1; // percent
+                SrcRef16 src{};
+                if (!readSrc(words, end, ip, src)) return;
+                handleSrc(src);
+            } break;
+
+            case Op16::Resize:
+                if (!need(2)) return;
+                ip += 2;
+                break;
+
+            case Op16::Transform:
+            case Op16::Opacity:
+                if (!need(1)) return;
+                ip += 1;
+                break;
+
+            case Op16::NoAlpha:
+            case Op16::Brighten:
+                break;
+
+            case Op16::MakeAlpha:
+                if (!need(2)) return;
+                ip += 2;
+                break;
+
+            case Op16::Invert:
+                if (!need(1)) return;
+                ip += 1;
+                break;
+
+            case Op16::Contrast:
+                if (!need(2)) return;
+                ip += 2;
+                break;
+
+            case Op16::Multiply:
+            case Op16::Screen:
+                if (!need(2)) return;
+                ip += 2;
+                break;
+
+            case Op16::Colorize:
+                if (!need(3)) return;
+                ip += 3;
+                break;
+
+            case Op16::Combine: {
+                if (!need(3)) return;
+                ip += 2; // skip w,h
+                uint32_t n = words[ip++];
+                for (uint32_t i = 0; i < n; ++i) {
+                    if (!need(2 + 3)) return;
+                    ip += 2; // x, y
+                    SrcRef16 src{};
+                    if (!readSrc(words, end, ip, src)) return;
+                    handleSrc(src);
+                }
+            } break;
+
+            default:
+                return;
+        }
+    }
+}
+
+inline boost::container::small_vector<uint32_t, 8> extractPipelineDependencies(const std::vector<Word>& words) {
+    boost::container::small_vector<uint32_t, 8> deps;
+    std::vector<std::pair<size_t, size_t>> visited;
+    extractPipelineDependencies(words, 0, words.size(), deps, visited);
+    return deps;
+}
+
+inline boost::container::small_vector<uint32_t, 8> extractPipelineDependencies(const boost::container::small_vector<Word, 32>& words) {
+    boost::container::small_vector<uint32_t, 8> deps;
+    std::vector<std::pair<size_t, size_t>> visited;
+    std::vector<Word> copy(words.begin(), words.end());
+    extractPipelineDependencies(copy, 0, copy.size(), deps, visited);
+    return deps;
+}
+
+} // namespace detail
+
+// Структура нехешированного пайплайна
+struct Pipeline {
+    std::vector<detail::Word> _Pipeline;
+
+    Pipeline() = default;
+
+    explicit Pipeline(const TexturePipelineProgram& program)
+        : _Pipeline(program.words().begin(), program.words().end())
+    {
+    }
+
+    Pipeline(TextureId texId) {
+        _Pipeline = {
+            static_cast<detail::Word>(detail::Op16::Base_Tex),
+            static_cast<detail::Word>(detail::SrcKind16::TexId),
+            static_cast<detail::Word>(texId & 0xFFFFu),
+            static_cast<detail::Word>((texId >> 16) & 0xFFFFu),
+            static_cast<detail::Word>(detail::Op16::End)
+        };
+    }
+};
+
+// Структура хешированного текстурного пайплайна
+struct HashedPipeline {
+    // Предвычисленный хеш
+    std::size_t _Hash;
+    boost::container::small_vector<detail::Word, 32> _Pipeline;
+
+    HashedPipeline() = default;
+    HashedPipeline(const Pipeline& pipeline) noexcept
+        : _Pipeline(pipeline._Pipeline.begin(), pipeline._Pipeline.end())
+    {
+        reComputeHash();
+    }
+
+    // Перевычисляет хеш
+    void reComputeHash() noexcept {
+        std::size_t hash = 14695981039346656037ull;
+        constexpr std::size_t prime = 1099511628211ull;
+
+        for(detail::Word w : _Pipeline) {
+            hash ^= static_cast<uint8_t>(w & 0xFF);
+            hash *= prime;
+            hash ^= static_cast<uint8_t>((w >> 8) & 0xFF);
+            hash *= prime;
+        }
+
+        _Hash = hash;
+    }
+
+    // Выдаёт список зависимых текстур, на основе которых строится эта
+    boost::container::small_vector<uint32_t, 8> getDependencedTextures() const {
+        return detail::extractPipelineDependencies(_Pipeline);
+    }
+
+    bool operator==(const HashedPipeline& obj) const noexcept {
+        return _Hash == obj._Hash && _Pipeline == obj._Pipeline;
+    }
+
+    bool operator<(const HashedPipeline& obj) const noexcept {
+        return _Hash < obj._Hash || (_Hash == obj._Hash && _Pipeline < obj._Pipeline);
+    }
+};
+
+struct StoredTexture {
+    uint16_t _Widht = 0;
+    uint16_t _Height = 0;
+    std::vector<uint32_t> _Pixels;
+
+    StoredTexture() = default;
+    StoredTexture(uint16_t w, uint16_t h, std::vector<uint32_t> pixels)
+        : _Widht(w), _Height(h), _Pixels(std::move(pixels))
+    {
+    }
+};
+
+
+// Пайплайновый текстурный атлас
+class PipelinedTextureAtlas {
+public:
+    using AtlasTextureId = uint32_t;
+    struct HostTextureView {
+        uint32_t width = 0;
+        uint32_t height = 0;
+        uint32_t rowPitchBytes = 0;
+        const uint8_t* pixelsRGBA8 = nullptr;
+    };
+
+private:
+    // Функтор хеша
+    struct HashedPipelineKeyHash {
+        std::size_t operator()(const HashedPipeline& k) const noexcept {
+            return k._Hash;
+        }
+    };
+
+    // Функтор равенства
+    struct HashedPipelineKeyEqual {
+        bool operator()(const HashedPipeline& a, const HashedPipeline& b) const noexcept {
+            return a._Pipeline == b._Pipeline;
+        }
+    };
+
+    // Текстурный атлас
+    TextureAtlas Super;
+    // Пустой пайплайн (указывающий на одну текстуру) ссылается на простой идентификатор (ResToAtlas)
+    std::unordered_map<HashedPipeline, AtlasTextureId, HashedPipelineKeyHash, HashedPipelineKeyEqual> _PipeToTexId;
+    // Загруженные текстуры
+    std::unordered_map<TextureId, StoredTexture> _ResToTexture;
+    std::unordered_map<AtlasTextureId, StoredTexture> _AtlasCpuTextures;
+    // Список зависимых пайплайнов от текстур (при изменении текстуры, нужно перерисовать пайплайны)
+    std::unordered_map<TextureId, boost::container::small_vector<HashedPipeline, 8>> _AddictedTextures;
+    // Изменённые простые текстуры (для последующего массового обновление пайплайнов)
+    std::vector<uint32_t> _ChangedTextures;
+    // Необходимые к созданию/обновлению пайплайны
+    std::vector<HashedPipeline> _ChangedPipelines;
+
+public:
+    PipelinedTextureAtlas(TextureAtlas&& tk);
+
+    uint32_t atlasSide() const {
+        return Super.atlasSide();
+    }
+
+    uint32_t atlasLayers() const {
+        return Super.atlasLayers();
+    }
+
+    uint32_t AtlasSide() const {
+        return atlasSide();
+    }
+
+    uint32_t AtlasLayers() const {
+        return atlasLayers();
+    }
+
+    // Должны всегда бронировать идентификатор, либо отдавать kOverflowId. При этом запись tex+pipeline остаётся
+    // Выдаёт стабильный идентификатор, привязанный к пайплайну
+    AtlasTextureId getByPipeline(const HashedPipeline& pipeline);
+
+    // Уведомить что текстура+pipeline более не используются (идентификатор будет освобождён)
+    // Освобождать можно при потере ресурсов
+    void freeByPipeline(const HashedPipeline& pipeline);
+
+    void updateTexture(uint32_t texId, const StoredTexture& texture);
+    void updateTexture(uint32_t texId, StoredTexture&& texture);
+
+    void freeTexture(uint32_t texId);
+
+    bool getHostTexture(TextureId texId, HostTextureView& out) const;
+
+    // Генерация текстуры пайплайна
+    StoredTexture _generatePipelineTexture(const HashedPipeline& pipeline);
+
+    // Обновляет пайплайны по необходимости
+    void flushNewPipelines();
+
+    TextureAtlas::DescriptorOut flushUploadsAndBarriers(VkCommandBuffer cmdBuffer);
+
+    void notifyGpuFinished();
+
+private:
+    std::optional<StoredTexture> tryCopyFirstDependencyTexture(const HashedPipeline& pipeline) const;
+
+    static StoredTexture makeSolidColorTexture(uint32_t rgba);
+};

Src/Client/Vulkan/AtlasPipeline/SharedStagingBuffer.hpp

@@ -0,0 +1,169 @@
+#pragma once
+
+#include <vulkan/vulkan.h>
+
+#include <cstdint>
+#include <optional>
+#include <stdexcept>
+#include <utility>
+
+/*
+  Межкадровый промежуточный буфер.
+  Для модели рендера Один за одним.
+  После окончания рендера кадра считается синхронизированным
+  и может заполняться по новой.
+*/
+
+class SharedStagingBuffer {
+public:
+  static constexpr VkDeviceSize kDefaultSize = 64ull * 1024ull * 1024ull;
+
+  SharedStagingBuffer(VkDevice device,
+                      VkPhysicalDevice physicalDevice,
+                      VkDeviceSize sizeBytes = kDefaultSize)
+      : device_(device),
+        physicalDevice_(physicalDevice),
+        size_(sizeBytes) {
+    if (!device_ || !physicalDevice_) {
+      throw std::runtime_error("SharedStagingBuffer: null device/physicalDevice");
+    }
+    if (size_ == 0) {
+      throw std::runtime_error("SharedStagingBuffer: size must be > 0");
+    }
+
+    VkBufferCreateInfo bi{
+      .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+      .pNext = nullptr,
+      .flags = 0,
+      .size = size_,
+      .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
+      .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+      .queueFamilyIndexCount = 0,
+      .pQueueFamilyIndices = nullptr
+    };
+
+    if (vkCreateBuffer(device_, &bi, nullptr, &buffer_) != VK_SUCCESS) {
+      throw std::runtime_error("SharedStagingBuffer: vkCreateBuffer failed");
+    }
+
+    VkMemoryRequirements mr{};
+    vkGetBufferMemoryRequirements(device_, buffer_, &mr);
+
+    VkMemoryAllocateInfo ai{};
+    ai.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
+    ai.allocationSize = mr.size;
+    ai.memoryTypeIndex = FindMemoryType_(mr.memoryTypeBits,
+                                         VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+                                             VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);
+
+    if (vkAllocateMemory(device_, &ai, nullptr, &memory_) != VK_SUCCESS) {
+      vkDestroyBuffer(device_, buffer_, nullptr);
+      buffer_ = VK_NULL_HANDLE;
+      throw std::runtime_error("SharedStagingBuffer: vkAllocateMemory failed");
+    }
+
+    vkBindBufferMemory(device_, buffer_, memory_, 0);
+
+    if (vkMapMemory(device_, memory_, 0, VK_WHOLE_SIZE, 0, &mapped_) != VK_SUCCESS) {
+      vkFreeMemory(device_, memory_, nullptr);
+      vkDestroyBuffer(device_, buffer_, nullptr);
+      buffer_ = VK_NULL_HANDLE;
+      memory_ = VK_NULL_HANDLE;
+      throw std::runtime_error("SharedStagingBuffer: vkMapMemory failed");
+    }
+  }
+
+  ~SharedStagingBuffer() { Destroy_(); }
+
+  SharedStagingBuffer(const SharedStagingBuffer&) = delete;
+  SharedStagingBuffer& operator=(const SharedStagingBuffer&) = delete;
+
+  SharedStagingBuffer(SharedStagingBuffer&& other) noexcept {
+    *this = std::move(other);
+  }
+
+  SharedStagingBuffer& operator=(SharedStagingBuffer&& other) noexcept {
+    if (this != &other) {
+      Destroy_();
+      device_ = other.device_;
+      physicalDevice_ = other.physicalDevice_;
+      buffer_ = other.buffer_;
+      memory_ = other.memory_;
+      mapped_ = other.mapped_;
+      size_ = other.size_;
+      offset_ = other.offset_;
+
+      other.device_ = VK_NULL_HANDLE;
+      other.physicalDevice_ = VK_NULL_HANDLE;
+      other.buffer_ = VK_NULL_HANDLE;
+      other.memory_ = VK_NULL_HANDLE;
+      other.mapped_ = nullptr;
+      other.size_ = 0;
+      other.offset_ = 0;
+    }
+    return *this;
+  }
+
+  VkBuffer Buffer() const { return buffer_; }
+  void* Mapped() const { return mapped_; }
+  VkDeviceSize Size() const { return size_; }
+
+  std::optional<VkDeviceSize> Allocate(VkDeviceSize bytes, VkDeviceSize alignment) {
+    VkDeviceSize off = Align_(offset_, alignment);
+    if (off + bytes > size_) {
+      return std::nullopt;
+    }
+    offset_ = off + bytes;
+    return off;
+  }
+
+  void Reset() { offset_ = 0; }
+
+private:
+  uint32_t FindMemoryType_(uint32_t typeBits, VkMemoryPropertyFlags properties) const {
+    VkPhysicalDeviceMemoryProperties mp{};
+    vkGetPhysicalDeviceMemoryProperties(physicalDevice_, &mp);
+    for (uint32_t i = 0; i < mp.memoryTypeCount; ++i) {
+      if ((typeBits & (1u << i)) &&
+          (mp.memoryTypes[i].propertyFlags & properties) == properties) {
+        return i;
+      }
+    }
+    throw std::runtime_error("SharedStagingBuffer: no suitable memory type");
+  }
+
+  static VkDeviceSize Align_(VkDeviceSize value, VkDeviceSize alignment) {
+    if (alignment == 0) return value;
+    return (value + alignment - 1) & ~(alignment - 1);
+  }
+
+  void Destroy_() {
+    if (device_ == VK_NULL_HANDLE) {
+      return;
+    }
+    if (mapped_) {
+      vkUnmapMemory(device_, memory_);
+      mapped_ = nullptr;
+    }
+    if (buffer_) {
+      vkDestroyBuffer(device_, buffer_, nullptr);
+      buffer_ = VK_NULL_HANDLE;
+    }
+    if (memory_) {
+      vkFreeMemory(device_, memory_, nullptr);
+      memory_ = VK_NULL_HANDLE;
+    }
+    size_ = 0;
+    offset_ = 0;
+    device_ = VK_NULL_HANDLE;
+    physicalDevice_ = VK_NULL_HANDLE;
+  }
+
+  VkDevice device_ = VK_NULL_HANDLE;
+  VkPhysicalDevice physicalDevice_ = VK_NULL_HANDLE;
+  VkBuffer buffer_ = VK_NULL_HANDLE;
+  VkDeviceMemory memory_ = VK_NULL_HANDLE;
+  void* mapped_ = nullptr;
+  VkDeviceSize size_ = 0;
+  VkDeviceSize offset_ = 0;
+};

Src/Client/Vulkan/AtlasPipeline/TextureAtlas.cpp

@@ -0,0 +1,477 @@
+#include "TextureAtlas.hpp"
+
+TextureAtlas::TextureAtlas(VkDevice device,
+                               VkPhysicalDevice physicalDevice,
+                               const Config& cfg,
+                               EventCallback cb,
+                               std::shared_ptr<SharedStagingBuffer> staging)
+    : Device_(device),
+      Phys_(physicalDevice),
+      Cfg_(cfg),
+      OnEvent_(std::move(cb)),
+      Staging_(std::move(staging)) {
+  if(!Device_ || !Phys_) {
+    throw std::runtime_error("TextureAtlas: device/physicalDevice == null");
+  }
+  _validateConfigOrThrow();
+
+  VkPhysicalDeviceProperties props{};
+  vkGetPhysicalDeviceProperties(Phys_, &props);
+  CopyOffsetAlignment_ = std::max<VkDeviceSize>(4, props.limits.optimalBufferCopyOffsetAlignment);
+
+  if(!Staging_) {
+    Staging_ = std::make_shared<SharedStagingBuffer>(Device_, Phys_, kStagingSizeBytes);
+  }
+  _validateStagingCapacityOrThrow();
+
+  _createEntriesBufferOrThrow();
+  _createAtlasOrThrow(Cfg_.InitialSide, 1);
+
+  EntriesCpu_.resize(Cfg_.MaxTextureId);
+  std::memset(EntriesCpu_.data(), 0, EntriesCpu_.size() * sizeof(Entry));
+  EntriesDirty_ = true;
+
+  Slots_.resize(Cfg_.MaxTextureId);
+  FreeIds_.reserve(Cfg_.MaxTextureId);
+  PendingInQueue_.assign(Cfg_.MaxTextureId, false);
+
+  if(Cfg_.ExternalSampler != VK_NULL_HANDLE) {
+    Sampler_ = Cfg_.ExternalSampler;
+    OwnsSampler_ = false;
+  } else {
+    _createSamplerOrThrow();
+    OwnsSampler_ = true;
+  }
+
+  _rebuildPackersFromPlacements();
+  Alive_ = true;
+}
+
+TextureAtlas::~TextureAtlas() { _shutdownNoThrow(); }
+
+TextureAtlas::TextureAtlas(TextureAtlas&& other) noexcept {
+  _moveFrom(std::move(other));
+}
+
+TextureAtlas& TextureAtlas::operator=(TextureAtlas&& other) noexcept {
+  if(this != &other) {
+    _shutdownNoThrow();
+    _moveFrom(std::move(other));
+  }
+  return *this;
+}
+
+void TextureAtlas::shutdown() {
+  _ensureAliveOrThrow();
+  _shutdownNoThrow();
+}
+
+TextureAtlas::TextureId TextureAtlas::registerTexture() {
+  _ensureAliveOrThrow();
+
+  TextureId id = kOverflowId;
+  if(!FreeIds_.empty()) {
+    id = FreeIds_.back();
+    FreeIds_.pop_back();
+  } else if(NextId_ < Cfg_.MaxTextureId) {
+    id = NextId_++;
+  } else {
+    return kOverflowId;
+  }
+
+  Slot& s = Slots_[id];
+  s = Slot{};
+  s.InUse = true;
+  s.StateValue = State::REGISTERED;
+  s.Generation = 1;
+
+  _setEntryInvalid(id, /*diagPending*/false, /*diagTooLarge*/false);
+  EntriesDirty_ = true;
+  return id;
+}
+
+void TextureAtlas::setTextureData(TextureId id,
+                                    uint32_t w,
+                                    uint32_t h,
+                                    const void* pixelsRGBA8,
+                                    uint32_t rowPitchBytes) {
+  _ensureAliveOrThrow();
+  if(id == kOverflowId) return;
+  _ensureRegisteredIdOrThrow(id);
+
+  if(w == 0 || h == 0) {
+    throw _inputError("setTextureData: w/h must be > 0");
+  }
+  if(w > Cfg_.MaxTextureSize || h > Cfg_.MaxTextureSize) {
+    _handleTooLarge(id);
+    throw _inputError("setTextureData: texture is TOO_LARGE (>2048)");
+  }
+  if(!pixelsRGBA8) {
+    throw _inputError("setTextureData: pixelsRGBA8 == null");
+  }
+
+  if(rowPitchBytes == 0) {
+    rowPitchBytes = w * 4;
+  }
+  if(rowPitchBytes < w * 4) {
+    throw _inputError("setTextureData: rowPitchBytes < w*4");
+  }
+
+  Slot& s = Slots_[id];
+
+  const bool sizeChanged = (s.HasCpuData && (s.W != w || s.H != h));
+  if(sizeChanged) {
+    _freePlacement(id);
+    _setEntryInvalid(id, /*diagPending*/true, /*diagTooLarge*/false);
+    EntriesDirty_ = true;
+  }
+
+  s.W = w;
+  s.H = h;
+
+  s.CpuPixels = static_cast<const uint8_t*>(pixelsRGBA8);
+  s.CpuRowPitchBytes = rowPitchBytes;
+  s.HasCpuData = true;
+  s.StateValue = State::PENDING_UPLOAD;
+  s.Generation++;
+
+  if(!sizeChanged && s.HasPlacement && s.StateWasValid) {
+    // keep entry valid
+  } else if(!s.HasPlacement) {
+    _setEntryInvalid(id, /*diagPending*/true, /*diagTooLarge*/false);
+    EntriesDirty_ = true;
+  }
+
+  _enqueuePending(id);
+
+  if(Repack_.Active && Repack_.Plan.count(id) != 0) {
+    _enqueueRepackPending(id);
+  }
+}
+
+void TextureAtlas::clearTextureData(TextureId id) {
+  _ensureAliveOrThrow();
+  if(id == kOverflowId) return;
+  _ensureRegisteredIdOrThrow(id);
+
+  Slot& s = Slots_[id];
+  s.CpuPixels = nullptr;
+  s.CpuRowPitchBytes = 0;
+  s.HasCpuData = false;
+
+  _freePlacement(id);
+  s.StateValue = State::REGISTERED;
+  s.StateWasValid = false;
+
+  _removeFromPending(id);
+  _removeFromRepackPending(id);
+
+  _setEntryInvalid(id, /*diagPending*/false, /*diagTooLarge*/false);
+  EntriesDirty_ = true;
+}
+
+void TextureAtlas::removeTexture(TextureId id) {
+  _ensureAliveOrThrow();
+  if(id == kOverflowId) return;
+  _ensureRegisteredIdOrThrow(id);
+
+  Slot& s = Slots_[id];
+
+  clearTextureData(id);
+
+  s.InUse = false;
+  s.StateValue = State::REMOVED;
+
+  FreeIds_.push_back(id);
+
+  _setEntryInvalid(id, /*diagPending*/false, /*diagTooLarge*/false);
+  EntriesDirty_ = true;
+}
+
+void TextureAtlas::requestFullRepack(RepackMode mode) {
+  _ensureAliveOrThrow();
+  Repack_.Requested = true;
+  Repack_.Mode = mode;
+}
+
+TextureAtlas::DescriptorOut TextureAtlas::flushUploadsAndBarriers(VkCommandBuffer cmdBuffer) {
+  _ensureAliveOrThrow();
+  if(cmdBuffer == VK_NULL_HANDLE) {
+    throw _inputError("flushUploadsAndBarriers: cmdBuffer == null");
+  }
+
+  if(Repack_.SwapReady) {
+    _swapToRepackedAtlas();
+  }
+  if(Repack_.Requested && !Repack_.Active) {
+    _startRepackIfPossible();
+  }
+
+  _processPendingLayerGrow(cmdBuffer);
+
+  bool willTouchEntries = EntriesDirty_;
+
+  auto collectQueue = [this](std::deque<TextureId>& queue,
+                             std::vector<bool>& inQueue,
+                             std::vector<TextureId>& out) {
+    while (!queue.empty()) {
+      TextureId id = queue.front();
+      queue.pop_front();
+      if(id == kOverflowId || id >= inQueue.size()) {
+        continue;
+      }
+      if(!inQueue[id]) {
+        continue;
+      }
+      inQueue[id] = false;
+      out.push_back(id);
+    }
+  };
+
+  std::vector<TextureId> pendingNow;
+  pendingNow.reserve(Pending_.size());
+  collectQueue(Pending_, PendingInQueue_, pendingNow);
+
+  std::vector<TextureId> repackPending;
+  if(Repack_.Active) {
+    if(Repack_.InPending.empty()) {
+      Repack_.InPending.assign(Cfg_.MaxTextureId, false);
+    }
+    collectQueue(Repack_.Pending, Repack_.InPending, repackPending);
+  }
+
+  auto processPlacement = [&](TextureId id, Slot& s) -> bool {
+    if(s.HasPlacement) return true;
+    const uint32_t wP = s.W + 2u * Cfg_.PaddingPx;
+    const uint32_t hP = s.H + 2u * Cfg_.PaddingPx;
+    if(!_tryPlaceWithGrow(id, wP, hP, cmdBuffer)) {
+      return false;
+    }
+    willTouchEntries = true;
+    return true;
+  };
+
+  bool outOfSpace = false;
+  for(TextureId id : pendingNow) {
+    if(id == kOverflowId) continue;
+    if(id >= Slots_.size()) continue;
+    Slot& s = Slots_[id];
+    if(!s.InUse || !s.HasCpuData) continue;
+    if(!processPlacement(id, s)) {
+      outOfSpace = true;
+      _enqueuePending(id);
+    }
+  }
+  if(outOfSpace) {
+    _emitEventOncePerFlush(AtlasEvent::AtlasOutOfSpace);
+  }
+
+  bool anyAtlasWrites = false;
+  bool anyRepackWrites = false;
+
+  auto uploadTextureIntoAtlas = [&](Slot& s,
+                                   const Placement& pp,
+                                   ImageRes& targetAtlas,
+                                   bool isRepackTarget) {
+    const uint32_t wP = pp.WP;
+    const uint32_t hP = pp.HP;
+    const VkDeviceSize bytes = static_cast<VkDeviceSize>(wP) * hP * 4u;
+    auto stagingOff = Staging_->Allocate(bytes, CopyOffsetAlignment_);
+    if(!stagingOff) {
+      _emitEventOncePerFlush(AtlasEvent::StagingOverflow);
+      return false;
+    }
+
+    uint8_t* dst = static_cast<uint8_t*>(Staging_->Mapped()) + *stagingOff;
+    if(!s.CpuPixels) {
+      return false;
+    }
+    _writePaddedRGBA8(dst, wP * 4u, s.W, s.H, Cfg_.PaddingPx,
+                      s.CpuPixels, s.CpuRowPitchBytes);
+
+    _ensureImageLayoutForTransferDst(cmdBuffer, targetAtlas,
+                                     isRepackTarget ? anyRepackWrites : anyAtlasWrites);
+
+    VkBufferImageCopy region{};
+    region.bufferOffset = *stagingOff;
+    region.bufferRowLength = wP;
+    region.bufferImageHeight = hP;
+    region.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+    region.imageSubresource.mipLevel = 0;
+    region.imageSubresource.baseArrayLayer = pp.Layer;
+    region.imageSubresource.layerCount = 1;
+    region.imageOffset = { static_cast<int32_t>(pp.X),
+                           static_cast<int32_t>(pp.Y), 0 };
+    region.imageExtent = { wP, hP, 1 };
+
+    vkCmdCopyBufferToImage(cmdBuffer, Staging_->Buffer(), targetAtlas.Image,
+                           VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &region);
+    return true;
+  };
+
+  for(TextureId id : pendingNow) {
+    if(id == kOverflowId) continue;
+    Slot& s = Slots_[id];
+    if(!s.InUse || !s.HasCpuData || !s.HasPlacement) continue;
+    if(!uploadTextureIntoAtlas(s, s.Place, Atlas_, false)) {
+      _enqueuePending(id);
+      continue;
+    }
+    s.StateValue = State::VALID;
+    s.StateWasValid = true;
+    _setEntryValid(id);
+    EntriesDirty_ = true;
+  }
+
+  if(Repack_.Active) {
+    for(TextureId id : repackPending) {
+      if(Repack_.Plan.count(id) == 0) continue;
+      Slot& s = Slots_[id];
+      if(!s.InUse || !s.HasCpuData) continue;
+      const PlannedPlacement& pp = Repack_.Plan[id];
+      Placement place{pp.X, pp.Y, pp.WP, pp.HP, pp.Layer};
+      if(!uploadTextureIntoAtlas(s, place, Repack_.Atlas, true)) {
+        _enqueueRepackPending(id);
+        continue;
+      }
+      Repack_.WroteSomethingThisFlush = true;
+    }
+  }
+
+  if(willTouchEntries || EntriesDirty_) {
+    const VkDeviceSize entriesBytes = static_cast<VkDeviceSize>(EntriesCpu_.size()) * sizeof(Entry);
+    auto off = Staging_->Allocate(entriesBytes, CopyOffsetAlignment_);
+    if(!off) {
+      _emitEventOncePerFlush(AtlasEvent::StagingOverflow);
+    } else {
+      std::memcpy(static_cast<uint8_t*>(Staging_->Mapped()) + *off,
+                  EntriesCpu_.data(),
+                  static_cast<size_t>(entriesBytes));
+
+      VkBufferCopy c{};
+      c.srcOffset = *off;
+      c.dstOffset = 0;
+      c.size = entriesBytes;
+      vkCmdCopyBuffer(cmdBuffer, Staging_->Buffer(), Entries_.Buffer, 1, &c);
+
+      VkBufferMemoryBarrier b{};
+      b.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
+      b.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+      b.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
+      b.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+      b.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+      b.buffer = Entries_.Buffer;
+      b.offset = 0;
+      b.size = VK_WHOLE_SIZE;
+
+      vkCmdPipelineBarrier(cmdBuffer,
+                           VK_PIPELINE_STAGE_TRANSFER_BIT,
+                           VK_PIPELINE_STAGE_VERTEX_SHADER_BIT |
+                               VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
+                               VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                           0, 0, nullptr, 1, &b, 0, nullptr);
+      EntriesDirty_ = false;
+    }
+  }
+
+  if(anyAtlasWrites) {
+    _transitionImage(cmdBuffer, Atlas_,
+                     VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+                     VK_ACCESS_TRANSFER_WRITE_BIT,
+                     VK_ACCESS_SHADER_READ_BIT,
+                     VK_PIPELINE_STAGE_TRANSFER_BIT,
+                     VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT);
+  } else if(Atlas_.Layout == VK_IMAGE_LAYOUT_UNDEFINED) {
+    _transitionImage(cmdBuffer, Atlas_,
+                     VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+                     0, VK_ACCESS_SHADER_READ_BIT,
+                     VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                     VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT);
+  }
+
+  if(anyRepackWrites) {
+    _transitionImage(cmdBuffer, Repack_.Atlas,
+                     VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+                     VK_ACCESS_TRANSFER_WRITE_BIT,
+                     VK_ACCESS_SHADER_READ_BIT,
+                     VK_PIPELINE_STAGE_TRANSFER_BIT,
+                     VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT);
+  }
+
+  if(Repack_.Active) {
+    if(Repack_.Pending.empty()) {
+      Repack_.WaitingGpuForReady = true;
+    }
+    Repack_.WroteSomethingThisFlush = false;
+  }
+
+  return _buildDescriptorOut();
+}
+
+void TextureAtlas::notifyGpuFinished() {
+  _ensureAliveOrThrow();
+
+  for(auto& img : DeferredImages_) {
+    _destroyImage(img);
+  }
+  DeferredImages_.clear();
+
+  if(Staging_) {
+    Staging_->Reset();
+  }
+  FlushEventMask_ = 0;
+
+  if(Repack_.Active && Repack_.WaitingGpuForReady && Repack_.Pending.empty()) {
+    Repack_.SwapReady = true;
+    Repack_.WaitingGpuForReady = false;
+  }
+}
+
+void TextureAtlas::_moveFrom(TextureAtlas&& other) noexcept {
+  Device_ = other.Device_;
+  Phys_ = other.Phys_;
+  Cfg_ = other.Cfg_;
+  OnEvent_ = std::move(other.OnEvent_);
+  Alive_ = other.Alive_;
+  CopyOffsetAlignment_ = other.CopyOffsetAlignment_;
+  Staging_ = std::move(other.Staging_);
+  Entries_ = other.Entries_;
+  Atlas_ = other.Atlas_;
+  Sampler_ = other.Sampler_;
+  OwnsSampler_ = other.OwnsSampler_;
+  EntriesCpu_ = std::move(other.EntriesCpu_);
+  EntriesDirty_ = other.EntriesDirty_;
+  Slots_ = std::move(other.Slots_);
+  FreeIds_ = std::move(other.FreeIds_);
+  NextId_ = other.NextId_;
+  Pending_ = std::move(other.Pending_);
+  PendingInQueue_ = std::move(other.PendingInQueue_);
+  Packers_ = std::move(other.Packers_);
+  DeferredImages_ = std::move(other.DeferredImages_);
+  FlushEventMask_ = other.FlushEventMask_;
+  GrewThisFlush_ = other.GrewThisFlush_;
+  Repack_ = std::move(other.Repack_);
+
+  other.Device_ = VK_NULL_HANDLE;
+  other.Phys_ = VK_NULL_HANDLE;
+  other.OnEvent_ = {};
+  other.Alive_ = false;
+  other.CopyOffsetAlignment_ = 0;
+  other.Staging_.reset();
+  other.Entries_ = {};
+  other.Atlas_ = {};
+  other.Sampler_ = VK_NULL_HANDLE;
+  other.OwnsSampler_ = false;
+  other.EntriesCpu_.clear();
+  other.EntriesDirty_ = false;
+  other.Slots_.clear();
+  other.FreeIds_.clear();
+  other.NextId_ = 0;
+  other.Pending_.clear();
+  other.PendingInQueue_.clear();
+  other.Packers_.clear();
+  other.DeferredImages_.clear();
+  other.FlushEventMask_ = 0;
+  other.GrewThisFlush_ = false;
+  other.Repack_ = RepackState{};
+}