Compare commits

...
6 Commits
Author SHA1 Message Date
swung0x48 56b31a9587 [Fix] (FastSTL): bump submodule for the erase(iterator) double-advance fix and add erase-while-iterating regression tests - the old semantics skipped one live element per erase and ran past end() when erasing the highest occupied bucket, sending the new mass pipeline-cache eviction sweeps off the bucket array (device crash on first eviction during world load: garbage handles fed to vkDestroyPipeline) 2026-07-27 23:50:25 -04:00
swung0x48 8a0a8a0274 [Fix] (DirectVulkan): harden the leak-fix round after adversarial review - pipeline memo now drops at every command-buffer boundary (a flush-loop-memoized pipeline could age out and be destroyed while its submission was in flight), mid-frame drains no longer rewind the arena or advance the cache-aging clocks in presenting apps (gated to every 8th drain since the last Present, so readback/fence-heavy frames neither churn conversions nor shrink the 1024-boundary retire window), render-pass eviction notifies the pipeline cache once per sweep batch instead of once per dying pass, descriptor pools use FREE_DESCRIPTOR_SET_BIT so a destroyed layout's cached sets are freed back and credited instead of abandoning pool slots (the live-layout age sweep that could orphan slots is removed - layout destruction is the sole purge path), and renderbuffer respecify parks the old backing for aged destruction instead of destroying it while possibly in flight 2026-07-27 22:51:26 -04:00
swung0x48 d076c29146 [Fix] (DirectVulkan): bound the vertex-input and sampler caches and sweep undeleted GL syncs - both caches age out entries idle >1024 frame boundaries (animated LOD bias no longer mints a VkSampler per float value, buffer/VAO churn no longer grows the vertex-input map for the whole session), and library teardown drains the live-sync registry exactly as glDeleteSync would since GL requires syncs to die with their context 2026-07-27 22:16:16 -04:00
swung0x48 930a607bdf [Fix] (DirectVulkan): make texture/renderbuffer GC reach every dead resource - name-deleted textures register via weak_from_this so first-sync-after-delete can no longer orphan a TextureResource, an orphan sweep makes GC authoritative over the resource map, dead-texture pruning moves to a frame-boundary gate (64 frames) so churn through clears/readbacks reclaims without draws, and dead renderbuffers age past frames-in-flight before their VkImage/view is destroyed instead of leaking until shutdown (or being freed while in flight) 2026-07-27 22:16:15 -04:00
swung0x48 34685b4bb0 [Fix] (DirectVulkan): age-based eviction for the content-addressed cache family - ProgramFactory entries (shader modules/layouts), PipelineFactory graphics pipelines, compute pipelines and per-layout descriptor-set tracking now retire after ~1024 idle frame boundaries (render-pass-manager sweep precedent), render-pass eviction purges pipelines hashed on the dying handle (closes a handle-recycling stale-pipeline hazard), and the program reflection cache is lifetime-id-keyed and cleared at EGL teardown - shader/program churn no longer grows Vulkan objects without bound 2026-07-27 22:05:25 -04:00
swung0x48 c540fb88ee [Fix] (DirectVulkan): drain frame transients on present-less paths - readback waits, suspended presentation, blocking sync waits and flush completion polls now run Present's per-frame drains (deferred buffer/texture releases, transient arena rewind, descriptor cursors, retired command buffers, conversion caches) whenever every submission is provably complete, so offscreen/minimized workloads stay bounded; never blocks, frames-in-flight overlap untouched 2026-07-27 21:45:21 -04:00
29 changed files with 995 additions and 42 deletions
+7
View File
@@ -14,6 +14,7 @@
#include <MG_State/EGLState/Core.h>
#include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
#include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
#include <MG_Impl/GLImpl/Sync/GL_Sync.h>
#include <atomic>
#include <mutex>
@@ -37,6 +38,12 @@ namespace MobileGL {
MGLOG_I("MobileGL closing...");
}
glslang::FinalizeProcess();
// GL syncs die with their contexts, and every context is gone by the
// time full teardown runs: drain the live-sync registry while the
// backend function table can still release the backend handles (and
// before a re-initialized library could pair them with the wrong
// backend's DeleteSync).
MG_Impl::GLImpl::DestroyAllSyncObjects();
MG_Backend::pActiveBackendObject.reset();
MG_State::pGLContext.reset();
MG_State::pEGLContext.reset();
@@ -469,6 +469,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// treat them as signaled/available with zero results from here on.
BumpRendererGeneration();
pVulkanRenderer.reset();
// The reflection cache is file-scope, not renderer-owned; without this the
// deleted programs' reflection strings survive full context teardown.
ClearProgramResourceCaches();
BackendObject::ReleaseEGLResources();
}
@@ -478,6 +481,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// treat them as signaled/available with zero results from here on.
BumpRendererGeneration();
pVulkanRenderer.reset();
// The reflection cache is file-scope, not renderer-owned; without this the
// deleted programs' reflection strings survive full context teardown.
ClearProgramResourceCaches();
}
const RendererInfo& BackendObject_DirectVulkan::GetRendererInfo() const {
@@ -61,6 +61,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
};
struct ProgramResourceCache {
// Lifetime id of the program the cached reflection belongs to. GL names are
// recycled (IndexGenerator hands freed indices straight back), and a
// recreated program's backendStateVersion restarts at the same small values,
// so the version alone can collide; the never-reused lifetime id makes the
// slot's ownership unambiguous.
Uint64 programLifetimeId = 0;
Uint32 backendStateVersion = 0;
Vector<StorageBlockResource> storageBlocks;
Vector<BufferVariableResource> bufferVariables;
@@ -82,6 +88,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint32 baseInstance = 0;
};
// Keyed by GL program name so the freed-name reuse in IndexGenerator bounds the
// map at the peak-simultaneous-program high-water mark; each slot's ownership is
// checked against the program's lifetime id before it is served (see
// GetProgramResourceCache). Cleared wholesale at EGL teardown via
// ClearProgramResourceCaches.
UnorderedMap<GLuint, ProgramResourceCache> g_programResourceCaches;
void ClearReadPixelsOutput(GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
@@ -142,13 +153,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
ProgramResourceCache& GetProgramResourceCache(const MG_State::GLState::ProgramObject& program) {
auto& cache = g_programResourceCaches[program.GetExternalIndex()];
const Uint64 programLifetimeId = program.GetLifetimeId();
const Uint32 backendStateVersion = program.GetBackendStateVersion();
if (cache.backendStateVersion == backendStateVersion &&
// The lifetime id must match too: a new program that reuses a deleted
// program's name and happens to land on the same backendStateVersion (both
// count from zero) would otherwise be served the dead program's reflection.
if (cache.programLifetimeId == programLifetimeId &&
cache.backendStateVersion == backendStateVersion &&
(!cache.storageBlocks.empty() || !cache.bufferVariables.empty())) {
return cache;
}
cache = {};
cache.programLifetimeId = programLifetimeId;
cache.backendStateVersion = backendStateVersion;
Vector<SpvReflectShaderModule> modules;
@@ -366,6 +383,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
} // namespace
void ClearProgramResourceCaches() {
// Called from EGL teardown while the backend's m_eglStateMutex is held; GL
// calls are serialized in this codebase (contexts migrate threads but never
// run concurrently), so no other thread can be inside the unsynchronized map.
// Live programs in another context self-heal: their entry rebuilds from the
// retained generated SPIR-V on the next resource query.
g_programResourceCaches.clear();
}
GLuint GetShaderStorageBlockIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
auto& cache = GetProgramResourceCache(program);
const auto it = std::find_if(cache.storageBlocks.begin(), cache.storageBlocks.end(),
@@ -23,6 +23,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint64 GetRendererGeneration();
void BumpRendererGeneration();
// Drops every cached program-resource reflection entry (CPU-side strings/vectors
// only, no Vulkan handles). Called at EGL teardown next to the renderer reset;
// safe because GL calls are serialized in this codebase, and any still-live
// program rebuilds its entry from the retained generated SPIR-V on demand.
void ClearProgramResourceCaches();
void ClearBufferfi(GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil);
void ClearBufferfv(GLenum buffer, GLint drawbuffer, const GLfloat* value);
void ClearBufferuiv(GLenum buffer, GLint drawbuffer, const GLuint* value);
@@ -264,7 +264,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
if (result != VK_SUCCESS) {
return result;
}
frame.retiredCommandBuffers.push_back(frame.commandBuffer);
// lastSubmitIndex was just written by the renderer for the submission
// that carried this command buffer.
frame.retiredCommandBuffers.push_back({frame.commandBuffer, frame.lastSubmitIndex});
frame.commandBuffer = replacement;
return VK_SUCCESS;
}
@@ -274,12 +276,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return;
}
if (m_device != VK_NULL_HANDLE && m_commandPool != VK_NULL_HANDLE) {
vkFreeCommandBuffers(m_device, m_commandPool, static_cast<Uint32>(frame.retiredCommandBuffers.size()),
frame.retiredCommandBuffers.data());
for (const auto& retired : frame.retiredCommandBuffers) {
vkFreeCommandBuffers(m_device, m_commandPool, 1, &retired.commandBuffer);
}
}
frame.retiredCommandBuffers.clear();
}
void FrameContext::FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex) {
if (m_device == VK_NULL_HANDLE || m_commandPool == VK_NULL_HANDLE) {
return;
}
for (auto& frame : m_frames) {
// Retired buffers are appended in submit order, so the completed
// ones form a prefix.
SizeT completedCount = 0;
while (completedCount < frame.retiredCommandBuffers.size() &&
frame.retiredCommandBuffers[completedCount].submitIndex <= completedSubmitIndex) {
vkFreeCommandBuffers(m_device, m_commandPool, 1,
&frame.retiredCommandBuffers[completedCount].commandBuffer);
++completedCount;
}
if (completedCount > 0) {
frame.retiredCommandBuffers.erase(frame.retiredCommandBuffers.begin(),
frame.retiredCommandBuffers.begin() + completedCount);
}
}
}
void FrameContext::FreeAllRetiredCommandBuffers() {
for (auto& frame : m_frames) {
FreeRetiredCommandBuffers(frame);
}
}
void FrameContext::AssertValidFrameIndex(Uint32 frameIndex) const {
MOBILEGL_ASSERT(frameIndex < m_frames.size(), "FrameContext index out of range");
}
@@ -40,6 +40,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
VkPresentInfoKHR presentInfo{VK_STRUCTURE_TYPE_PRESENT_INFO_KHR};
};
// A command buffer submitted mid-frame (FlushPendingCommands), tagged
// with the submit-tracker index it was submitted under so it can be
// freed as soon as that submission is observed complete - without
// waiting for the slot's fence to be waited again (present-less flush
// loops never wait it).
struct RetiredCommandBuffer {
VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
Uint64 submitIndex = 0;
};
struct FrameData {
VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
VkSemaphore imageAvailableSemaphore = VK_NULL_HANDLE;
@@ -47,10 +57,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Bool isCommandRecording = false;
Bool hasCommandBufferRecorded = false;
Bool imageAvailableSemaphoreConsumed = false;
// Command buffers submitted mid-frame (FlushPendingCommands) whose
// execution is only known complete once this slot's fence has been
// waited again; freed at that point.
Vector<VkCommandBuffer> retiredCommandBuffers;
// Command buffers submitted mid-frame (FlushPendingCommands),
// appended in submit order; freed once their submission is known
// complete (fence wait or completion poll).
Vector<RetiredCommandBuffer> retiredCommandBuffers;
// Submit-tracker index of this slot's most recent queue submission
// (written by the renderer at submit time).
Uint64 lastSubmitIndex = 0;
@@ -79,9 +89,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// Parks the current (already ended and submitted) command buffer on the
// slot's retired list and installs a freshly allocated one, so recording
// can restart while the submitted buffer is still executing. Retired
// buffers are freed after the slot's fence is next waited.
// buffers are freed after the slot's fence is next waited, or as soon
// as their submission is observed complete.
VkResult RetireCurrentCommandBuffer();
// Frees every retired command buffer whose tagged submission index is
// known complete. Driven by the renderer's submit tracker on completion
// events (fence waits and non-blocking polls), so present-less flush
// loops reclaim their buffers without any extra wait.
void FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex);
// Frees every slot's retired command buffers. Only valid when the
// caller has proven every queue submission complete.
void FreeAllRetiredCommandBuffers();
Uint32 GetCurrentFrameIndex() const;
Uint32 GetFrameCount() const;
@@ -8,6 +8,7 @@
#include "PipelineFactory.h"
#include <algorithm>
namespace MobileGL::MG_Backend::DirectVulkan {
static const char* PrimitiveTopologyToString(VkPrimitiveTopology topology) {
@@ -243,23 +244,108 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const HashType hash = ComputeHash(payload);
auto it = m_cache.find(hash);
if (it != m_cache.end()) {
return it->second;
it->second.lastUsedFrame = m_frameCounter;
return it->second.pipeline;
}
VkPipeline pipeline = CreatePipeline(payload);
m_cache.emplace(hash, pipeline);
m_cache.emplace(hash, PipelineCacheEntry{pipeline, payload.programHash, payload.renderPass,
m_frameCounter});
return pipeline;
}
void PipelineFactory::DestroyAll() {
for (auto& pair : m_cache) {
if (pair.second != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, pair.second, nullptr);
if (pair.second.pipeline != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, pair.second.pipeline, nullptr);
}
}
m_cache.clear();
}
Uint32 PipelineFactory::OnFrameBoundary() {
++m_frameCounter;
// Sweep cadence and retire age mirror VkRenderPassManager::OnPresent: an entry
// idle for more than kRetireAgeFrames frame boundaries cannot be referenced by
// any in-flight command buffer (frames-in-flight <= MOBILEGL_MAGMA_FRAMESINFLIGHT),
// so immediate vkDestroyPipeline is safe. The caller must drop its "last
// pipeline" memo when this returns non-zero: the memo can return a cached
// handle without touching this cache, so an evicted pipeline may still be
// memoized (present-less flush loops never reset the memo per frame).
constexpr Uint64 kSweepInterval = 256;
constexpr Uint64 kRetireAgeFrames = 1024;
if ((m_frameCounter % kSweepInterval) != 0) {
return 0;
}
Uint32 evicted = 0;
for (auto it = m_cache.begin(); it != m_cache.end();) {
if (m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
if (it->second.pipeline != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
}
it = m_cache.erase(it);
++evicted;
} else {
++it;
}
}
if (evicted > 0) {
MGLOG_D("PipelineFactory::OnFrameBoundary: evicted %u idle pipelines (%zu remain)", evicted,
m_cache.size());
}
return evicted;
}
Uint32 PipelineFactory::EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses) {
if (renderPasses.empty() || m_cache.empty()) {
return 0;
}
// Sorted-batch membership test keeps a mass eviction (shader-pack switch,
// dimension exit) at one O(cache * log batch) scan instead of one full scan
// per dying pass.
Vector<VkRenderPass> sortedPasses = renderPasses;
std::sort(sortedPasses.begin(), sortedPasses.end());
Uint32 evicted = 0;
for (auto it = m_cache.begin(); it != m_cache.end();) {
if (std::binary_search(sortedPasses.begin(), sortedPasses.end(), it->second.renderPass)) {
if (it->second.pipeline != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
}
it = m_cache.erase(it);
++evicted;
} else {
++it;
}
}
if (evicted > 0) {
MGLOG_D("PipelineFactory::EvictByRenderPasses: evicted %u pipelines for %zu destroyed render passes",
evicted, sortedPasses.size());
}
return evicted;
}
Uint32 PipelineFactory::EvictByProgramHash(HashType programHash) {
Uint32 evicted = 0;
for (auto it = m_cache.begin(); it != m_cache.end();) {
if (it->second.programHash == programHash) {
if (it->second.pipeline != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
}
it = m_cache.erase(it);
++evicted;
} else {
++it;
}
}
if (evicted > 0) {
MGLOG_D("PipelineFactory::EvictByProgramHash: evicted %u pipelines for program hash 0x%llx",
evicted, static_cast<unsigned long long>(programHash));
}
return evicted;
}
VkPipeline PipelineFactory::CreatePipeline(const PipelineCreatePayload& payload) const {
MOBILEGL_ASSERT(payload.stages != nullptr && !payload.stages->empty(), "PipelineFactory: stages are empty");
MOBILEGL_ASSERT(payload.vertexInputState != nullptr, "PipelineFactory: vertexInputState is null");
@@ -65,6 +65,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
VkPipeline GetOrCreatePipeline(const PipelineCreatePayload& payload);
void DestroyAll();
// Frame boundary hook: ages the pipeline cache and destroys long-unused entries
// (their command buffers retired many frames ago), mirroring
// VkRenderPassManager::OnPresent's sweep. Returns the number of pipelines
// destroyed so the caller can drop any memoized VkPipeline handle.
Uint32 OnFrameBoundary();
// Destroys every cached pipeline hashed on one of `renderPasses`. Only safe
// when the caller guarantees GPU idleness for them - the render-pass manager
// calls this (via the renderer) for passes its own >1024-boundary-idle sweep
// just evicted, and a pipeline hashed on those handles is only ever bound by
// draws that also hit the render-pass entries. Also closes the handle-recycling
// hazard: a recycled VkRenderPass value must never serve a stale pipeline.
// Batched: one cache scan regardless of how many passes died in the sweep.
// Returns the number destroyed (callers invalidate memos when non-zero).
Uint32 EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses);
// Destroys every cached pipeline built from the program with content hash
// `programHash`. Called from the ProgramFactory eviction path, which proves the
// same >1024-boundary idleness (the program's pipelines are only bound by draws
// that stamp its factory entry). Returns the number destroyed.
Uint32 EvictByProgramHash(HashType programHash);
// Driver quirk: suppress depth writes on accumulation-blended pipelines. Multi-pass
// depth-equality rendering (a blended prepass writes depth that later passes re-test
// with an equality-inclusive compare on the re-rasterized geometry) requires
@@ -87,12 +107,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
static Bool ShouldSuppressDepthWrite(const PipelineCreatePayload& payload);
private:
struct PipelineCacheEntry {
VkPipeline pipeline = VK_NULL_HANDLE;
// The hashed inputs the eviction paths key on: programHash ties the entry to
// its ProgramFactory entry, renderPass records the exact handle the hash
// folded in (the hash is one-way, so targeted eviction needs them verbatim).
HashType programHash = 0;
VkRenderPass renderPass = VK_NULL_HANDLE;
// Frame-boundary counter value of the last GetOrCreatePipeline hit; drives
// cache eviction (see OnFrameBoundary).
Uint64 lastUsedFrame = 0;
};
VkPipeline CreatePipeline(const PipelineCreatePayload& payload) const;
VkDevice m_device = VK_NULL_HANDLE;
const VulkanRendererConfig& m_config;
VkPipelineCache m_pipelineCache = VK_NULL_HANDLE;
UnorderedMap<HashType, VkPipeline> m_cache;
UnorderedMap<HashType, PipelineCacheEntry> m_cache;
// Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
Uint64 m_frameCounter = 0;
static inline XXH64_state_t* m_hashState = XXH64_createState();
static inline Bool s_suppressBlendedDepthWrite = false;
};
@@ -1950,11 +1950,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
auto it = m_cache.find(hash);
if (it != m_cache.end()) {
// Every draw/dispatch funnels through this lookup (the renderer memos only
// skip re-hashing, never the factory lookup), so an actively-used entry is
// stamped at least once per frame boundary and can never be aged out while
// any in-flight command buffer still references it.
it->second.lastUsedFrame = m_frameCounter;
return it->second;
}
auto& entry = m_cache[hash];
entry.hash = hash;
entry.lastUsedFrame = m_frameCounter;
auto& shaders = program.GetAttachedShaders();
auto& spirv = program.GetGeneratedSpirv();
Vector<Vector<Uint>> moduleSpirvs(spirv.size());
@@ -2068,4 +2074,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return entry;
}
void ProgramFactory::OnFrameBoundary() {
++m_frameCounter;
// Sweep cadence and retire age mirror VkRenderPassManager::OnPresent: an entry
// idle for more than kRetireAgeFrames frame boundaries cannot be referenced by
// any in-flight command buffer (frames-in-flight <= MOBILEGL_MAGMA_FRAMESINFLIGHT),
// so its shader modules and layouts are destroyed immediately - no deferred-
// destroy machinery needed. Eviction is content-based, never tied to
// glDeleteProgram: the cache is content-hash-shared across GL programs, so a
// delete-driven erase could free an entry another live program still resolves.
// An evicted entry self-heals - the frontend program keeps its generated
// SPIR-V, so the next GetOrCreateProgram rebuilds it (this also covers the
// renderer's internal blit/depth-mipmap programs).
constexpr Uint64 kSweepInterval = 256;
constexpr Uint64 kRetireAgeFrames = 1024;
if ((m_frameCounter % kSweepInterval) != 0) {
return;
}
for (auto it = m_cache.begin(); it != m_cache.end();) {
if (m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
const HashType hash = it->first;
const VkDescriptorSetLayout descriptorSetLayout = it->second.descriptorSetLayout;
MGLOG_D("ProgramFactory::OnFrameBoundary: evicting idle program entry hash=0x%llx",
static_cast<unsigned long long>(hash));
// erase runs ~VkProgramObject (modules/layouts destroyed); notify after
// so an observer never observes a half-destroyed entry through a lookup.
// Observers only need the handle values to purge their keyed caches.
it = m_cache.erase(it);
if (m_evictionObserver != nullptr) {
m_evictionObserver->OnProgramEvicted(hash, descriptorSetLayout);
}
} else {
++it;
}
}
}
} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -88,6 +88,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// gl_FragDepth); shader-computed depth is immune to the cross-pipeline
// position-invariance quirk (see PipelineFactory::ShouldSuppressDepthWrite).
Bool fragmentReplacesDepth = false;
// Frame-boundary counter value of the last GetOrCreateProgram hit; drives
// cache eviction (see OnFrameBoundary).
Uint64 lastUsedFrame = 0;
static inline VkDevice s_device = VK_NULL_HANDLE;
@@ -124,6 +127,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
producerOutputComponentCount = other.producerOutputComponentCount;
fragmentInputComponentCount = other.fragmentInputComponentCount;
fragmentReplacesDepth = other.fragmentReplacesDepth;
lastUsedFrame = other.lastUsedFrame;
other.hash = 0;
other.descriptorSetLayout = VK_NULL_HANDLE;
other.pipelineLayout = VK_NULL_HANDLE;
@@ -135,6 +139,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
other.producerOutputComponentCount = 0;
other.fragmentInputComponentCount = 0;
other.fragmentReplacesDepth = false;
other.lastUsedFrame = 0;
}
VkProgramObject& operator=(VkProgramObject&& other) noexcept {
if (this == &other) {
@@ -170,6 +175,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
producerOutputComponentCount = other.producerOutputComponentCount;
fragmentInputComponentCount = other.fragmentInputComponentCount;
fragmentReplacesDepth = other.fragmentReplacesDepth;
lastUsedFrame = other.lastUsedFrame;
other.hash = 0;
other.descriptorSetLayout = VK_NULL_HANDLE;
other.pipelineLayout = VK_NULL_HANDLE;
@@ -181,6 +187,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
other.producerOutputComponentCount = 0;
other.fragmentInputComponentCount = 0;
other.fragmentReplacesDepth = false;
other.lastUsedFrame = 0;
return *this;
}
@@ -210,6 +217,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
};
// Notified when the OnFrameBoundary sweep destroys an aged-out cache entry,
// carrying the entry's content hash and the VkDescriptorSetLayout it owned.
// Dependent caches (compute pipelines, PipelineFactory entries, UniformManager's
// per-layout descriptor sets) must purge in the same step: after vkDestroy the
// layout handle value may be recycled for an unrelated layout, and the program
// hash may be re-inserted by a later rebuild of the same content.
class IEvictionObserver {
public:
virtual ~IEvictionObserver() = default;
virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
};
explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
Bool shaderDrawParametersEnabled = false,
Bool unformattedFloatStorageImagesEnabled = false)
@@ -225,6 +244,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const VkProgramObject& GetOrCreateProgram(
const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
// Observer may be null (no notifications). Not owned.
void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
// Frame boundary hook: ages the program cache and evicts long-unused entries
// (their command buffers retired many frames ago), mirroring
// VkRenderPassManager::OnPresent's sweep.
void OnFrameBoundary();
static VkShaderStageFlagBits ToVkStage(ShaderStage stage);
static VkFormat ConvertSpirvImageFormatToVkFormat(SpvImageFormat format);
static SamplerNumericDomain UniformTypeToSamplerNumericDomain(GLenum glType);
@@ -266,6 +292,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
Bool m_unformattedFloatStorageImagesEnabled = false;
mutable ProgramLookupCache m_lastLookup;
// Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
Uint64 m_frameCounter = 0;
IEvictionObserver* m_evictionObserver = nullptr;
static inline XXH64_state_t* m_hashState = XXH64_createState();
};
} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -211,6 +211,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
}
void UniformManager::OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout) {
SizeT purgedSets = 0;
for (auto& frame : m_frames) {
const auto it = frame.descriptorSetCacheByLayout.find(descriptorSetLayout);
if (it == frame.descriptorSetCacheByLayout.end()) {
continue;
}
// Free the sets back to their pools and credit the bucket accounting, so
// program churn recycles pool capacity instead of abandoning the slots.
// GPU-safe: the layout only dies after >1024 idle frame boundaries, so no
// in-flight command buffer references these sets.
for (const auto& cached : it->second.sets) {
if (cached.set == VK_NULL_HANDLE) {
continue;
}
vkFreeDescriptorSets(m_device, cached.pool, 1, &cached.set);
const auto bucket = std::find_if(
frame.descriptorPools.begin(), frame.descriptorPools.end(),
[&cached](const DescriptorPoolBucket& candidate) { return candidate.handle == cached.pool; });
if (bucket != frame.descriptorPools.end() && bucket->allocatedSets > 0) {
--bucket->allocatedSets;
}
}
purgedSets += it->second.sets.size();
frame.descriptorSetCacheByLayout.erase(it);
}
if (purgedSets > 0) {
// The per-draw reuse memo folds the layout handle into its signature; drop
// it so a recycled handle value cannot revive a purged set mid-frame.
m_hasLastDescriptor = false;
MGLOG_D("UniformDescriptorBinder: freed %zu descriptor sets for destroyed layout", purgedSets);
}
}
Bool UniformManager::ResolveSamplerDescriptor(VkCommandBuffer commandBuffer,
const MG_State::GLState::ProgramObject& program,
const ProgramFactory::VkProgramObject& programObj,
@@ -911,6 +945,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
VkDescriptorPoolCreateInfo poolInfo{};
poolInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
// FREE_DESCRIPTOR_SET_BIT lets a destroyed layout's cached sets be freed back
// (OnDescriptorSetLayoutDestroyed) so program churn recycles pool capacity.
// The cost is on set allocation only, which happens when a layout's per-frame
// cache grows - never on the per-draw reuse path.
poolInfo.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT;
poolInfo.maxSets = maxSets;
poolInfo.poolSizeCount = static_cast<Uint32>(std::size(poolSizes));
poolInfo.pPoolSizes = poolSizes;
@@ -990,7 +1029,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
auto& frame = m_frames[frameIndex];
auto& cache = frame.descriptorSetCacheByLayout[programObj.descriptorSetLayout];
if (cache.cursor < cache.sets.size()) {
outDescriptorSet = cache.sets[cache.cursor++];
outDescriptorSet = cache.sets[cache.cursor++].set;
} else {
VkResult allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
if (allocResult == VK_ERROR_OUT_OF_POOL_MEMORY || allocResult == VK_ERROR_FRAGMENTED_POOL) {
@@ -1004,7 +1043,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return allocResult;
}
cache.sets.push_back(outDescriptorSet);
// The successful allocation came from the bucket the alloc helper left
// active; record it so a layout-destroyed purge can free the set back.
cache.sets.push_back({outDescriptorSet, frame.descriptorPools[frame.activeDescriptorPoolIndex].handle});
++cache.cursor;
MGLOG_D("UniformDescriptorBinder: cached descriptor set count for frame=%u grew to %zu", frameIndex,
cache.sets.size());
@@ -39,6 +39,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
void Shutdown();
void BeginFrame(Uint32 frameIndex);
// A ProgramFactory eviction just destroyed this layout: purge every frame
// slot's cached descriptor sets for it, so a recycled handle value can never
// stale-hit sets written for the dead layout's bindings. The sets are
// vkFreeDescriptorSets'd back to their pools (created with
// FREE_DESCRIPTOR_SET_BIT) and the pool accounting is credited, so program
// churn recycles pool capacity instead of abandoning it. GPU-safe: the layout
// only dies after >1024 idle frame boundaries, so no in-flight command buffer
// references its sets. This is the only eviction path for the per-layout
// caches - a live layout's entry must never be purged (its sets would be
// unreachable pool slots), so there is deliberately no age-based sweep here.
void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
Bool CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
const ProgramFactory::VkProgramObject& programObj,
Vector<MG_State::GLState::ITextureObject*>& outTextures);
@@ -65,8 +76,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint32 allocatedSets = 0;
};
// A cached descriptor set together with the pool it was allocated from, so a
// layout-destroyed purge can vkFreeDescriptorSets it back and credit the
// owning bucket's accounting.
struct CachedDescriptorSet {
VkDescriptorSet set = VK_NULL_HANDLE;
VkDescriptorPool pool = VK_NULL_HANDLE;
};
struct DescriptorSetCacheEntry {
Vector<VkDescriptorSet> sets;
Vector<CachedDescriptorSet> sets;
Uint32 cursor = 0;
};
@@ -32,6 +32,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsBgra, sizeof(attr.IsBgra)));
XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Divisor, sizeof(attr.Divisor)));
// The buffer's heap address is an identity component of the key: a freed
// buffer's reused address can alias an old cache entry, but only under a
// byte-identical attribute layout - and the entry payload is a pure function
// of the hashed inputs, with the draw path re-resolving bindingBufferKeys
// against the live VAO attribute pointers, so an aliased hit returns exactly
// what a rebuild would. Address drift only grows the map; the OnFrameBoundary
// aging sweep bounds that.
const SizeT bufferKey = reinterpret_cast<SizeT>(attr.Buffer.get());
XXHASH_VERIFY(XXH64_update(m_hashState, &bufferKey, sizeof(bufferKey)));
}
@@ -58,6 +65,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const MG_State::GLState::VertexArrayObject& vao, HashType hash) {
auto it = m_cache.find(hash);
if (it != m_cache.end()) {
it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
return it->second;
}
@@ -166,6 +174,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
auto& entry = m_cache[hash];
entry.hash = hash;
entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
entry.bindings = builder.GetBindings();
entry.attributes = builder.GetAttributes();
entry.bindingBufferKeys = std::move(bindingBufferKeys);
@@ -180,6 +189,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return entry;
}
void VertexInputStateFactory::OnFrameBoundary() {
++m_frameBoundaryCounter;
// Sweep occasionally; evict entries whose last hit is far in the past.
// Erasure happens only here, never mid-frame: the draw path holds a
// reference into the current entry across its setup, and unordered_map
// erase would invalidate it. Entries are CPU-side only, so no GPU-idle
// proof is needed; an evicted entry that is used again is simply rebuilt
// from the VAO state (same hash, same content).
constexpr Uint64 kSweepInterval = 256;
constexpr Uint64 kRetireAgeBoundaries = 1024;
if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
return;
}
for (auto it = m_cache.begin(); it != m_cache.end();) {
if (m_frameBoundaryCounter - it->second.lastUsedFrameBoundary > kRetireAgeBoundaries) {
it = m_cache.erase(it);
} else {
++it;
}
}
}
VkFormat VertexInputStateFactory::ToVkVertexFormat(DataType type, Int size, Bool normalized, Bool isInteger,
Bool isBgra) {
if (isBgra) {
@@ -27,6 +27,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
struct BackendVertexInputState {
HashType hash = 0;
// Frame boundary of the last cache hit; entries idle past the
// OnFrameBoundary retirement age are evicted (CPU heap only).
Uint64 lastUsedFrameBoundary = 0;
Vector<VkVertexInputBindingDescription> bindings;
Vector<VkVertexInputAttributeDescription> attributes;
Vector<SizeT> bindingBufferKeys;
@@ -55,6 +58,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const BackendVertexInputState& GetOrCreateVertexInputState(
const MG_State::GLState::VertexArrayObject& vao, HashType hash);
const BackendVertexInputState& GetOrCreateVertexInputState(const MG_State::GLState::VertexArrayObject& vao);
// Frame boundary hook: ages the cache and evicts entries not hit for many
// frames. The key mixes buffer heap addresses, so buffer/VAO churn keeps
// minting fresh keys; without eviction the map grows for the whole session.
// Entries hold no Vulkan handles (pipeline creation copies the descriptions)
// and the draw path's entry reference never spans a frame boundary, so
// eviction here needs no GPU-idle proof. Self-gated: one counter bump and
// compare except on sweep boundaries.
void OnFrameBoundary();
static SizeT GetComponentSize(DataType type);
// Tightly-packed byte size of one vertex element for this attribute: componentSize * size for
// normal types, and 4 (one packed word) for the 2_10_10_10 types and GL_BGRA. Returns 0 for
@@ -70,6 +81,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const VulkanRendererConfig& m_config;
VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
UnorderedMap<HashType, BackendVertexInputState> m_cache;
// Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
Uint64 m_frameBoundaryCounter = 0;
static inline XXH64_state_t* m_hashState = XXH64_createState();
};
} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -141,6 +141,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
m_transientUploadArena.BeginFrame(frameIndex);
}
void VkBufferManager::CollectAllDeferredReleases() {
for (Uint32 frameIndex = 0; frameIndex < m_deferredBufferReleases.size(); ++frameIndex) {
CollectDeferredReleases(frameIndex);
}
for (Uint32 frameIndex = 0; frameIndex < m_transientUploadArena.GetFrameCount(); ++frameIndex) {
m_transientUploadArena.CollectDeferredReleases(frameIndex);
}
}
void VkBufferManager::NotifyDeviceIdle() {
// Everything submitted so far has completed. Work recorded for the
// current frame has not been submitted yet, so the current serial
@@ -77,6 +77,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// Recreate all per-frame transient arenas
Bool RecreateTransientArenas(Uint32 frameCount);
void BeginFrame(Uint32 frameIndex);
// Drains every frame slot's deferred buffer/resource releases (and the
// transient arena's parked superseded blocks). Only valid when the
// caller has proven every queue submission complete; used by the
// present-less frame-boundary drain.
void CollectAllDeferredReleases();
// All previously submitted GPU work has completed (vkDeviceWaitIdle).
void NotifyDeviceIdle();
// A frame slot's submission fence has been waited: every serial up to
@@ -180,6 +180,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
sampleCount = VK_SAMPLE_COUNT_1_BIT;
internalFormat = TextureInternalFormat::Unknown;
samples = 0;
deadSinceFrame = kNeverObservedDead;
}
VkRenderPassManager::VkRenderPassManager(VkDevice device,
@@ -206,6 +207,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
resource.Destroy(m_device, m_allocator);
}
m_renderbufferResources.clear();
CollectDeferredRenderbufferReleases(/*destroyAll=*/true); // caller guarantees device idle
m_pendingRenderbufferClears.clear();
RenderPassEntry::s_textureResourcesScratch.clear();
s_activeRenderPass = {};
@@ -213,22 +215,75 @@ namespace MobileGL::MG_Backend::DirectVulkan {
m_rpFastValid = false;
}
Uint64 VkRenderPassManager::RetireAgeFrames() const {
// MaxFramesInFlight + 2 covers the frame ring plus one boundary for the
// recording-to-submit gap and one because OnPresent runs ahead of Present's
// fence wait; the floor of 8 keeps a margin over the default ring of 3 while
// still releasing multi-MB attachment memory promptly (the render-pass cache's
// 1024-frame retirement would pin it for no additional safety).
return std::max<Uint64>(8, static_cast<Uint64>(m_config.MaxFramesInFlight) + 2);
}
void VkRenderPassManager::DeferRenderbufferBackingRelease(RenderbufferResource& resource) {
// The superseded backing may still be referenced by in-flight command buffers
// (glRenderbufferStorage can respecify a renderbuffer drawn this very frame),
// so it is parked and destroyed only after RetireAgeFrames() boundaries.
if (resource.image == VK_NULL_HANDLE && resource.view == VK_NULL_HANDLE) {
return;
}
m_deferredRenderbufferReleases.push_back({resource.image, resource.allocation, resource.view, m_frameCounter});
resource.image = VK_NULL_HANDLE;
resource.allocation = nullptr;
resource.view = VK_NULL_HANDLE;
}
void VkRenderPassManager::CollectDeferredRenderbufferReleases(Bool destroyAll) {
if (m_deferredRenderbufferReleases.empty()) {
return;
}
const Uint64 retireAgeFrames = RetireAgeFrames();
std::erase_if(m_deferredRenderbufferReleases, [&](DeferredRenderbufferRelease& release) {
if (!destroyAll && m_frameCounter - release.deferredAtFrame < retireAgeFrames) {
return false;
}
if (release.view != VK_NULL_HANDLE) {
vkDestroyImageView(m_device, release.view, nullptr);
}
if (release.image != VK_NULL_HANDLE) {
vmaDestroyImage(m_allocator, release.image, release.allocation);
}
return true;
});
}
void VkRenderPassManager::CollectRenderbufferGarbage() {
Vector<MG_State::GLState::RenderbufferObject*> deadRenderbuffers;
deadRenderbuffers.reserve(m_renderbufferResources.size());
for (auto& [renderbuffer, resource] : m_renderbufferResources) {
// Two-phase reclamation: a dead renderbuffer's VkImage may still be referenced by
// command buffers submitted up to frames-in-flight frames ago (it was legally
// attached and drawn right up to its deletion), so the first observation of an
// expired weak reference only stamps the current frame counter; Destroy runs once
// enough frame boundaries have passed that the stamping frame's submission fence
// has provably been waited (see RetireAgeFrames).
const Uint64 retireAgeFrames = RetireAgeFrames();
for (auto it = m_renderbufferResources.begin(); it != m_renderbufferResources.end();) {
auto& resource = it->second;
const auto liveRenderbuffer = resource.renderbuffer.lock();
if (!liveRenderbuffer || liveRenderbuffer.get() != renderbuffer) {
deadRenderbuffers.emplace_back(renderbuffer);
if (liveRenderbuffer && liveRenderbuffer.get() == it->first) {
resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
++it;
continue;
}
if (resource.deadSinceFrame == RenderbufferResource::kNeverObservedDead) {
resource.deadSinceFrame = m_frameCounter;
++it;
continue;
}
for (auto* renderbuffer : deadRenderbuffers) {
auto resourceIt = m_renderbufferResources.find(renderbuffer);
if (resourceIt != m_renderbufferResources.end()) {
resourceIt->second.Destroy(m_device, m_allocator);
m_renderbufferResources.erase(resourceIt);
if (m_frameCounter - resource.deadSinceFrame < retireAgeFrames) {
++it;
continue;
}
m_pendingRenderbufferClears.erase(renderbuffer);
m_pendingRenderbufferClears.erase(it->first);
resource.Destroy(m_device, m_allocator);
it = m_renderbufferResources.erase(it);
}
}
@@ -270,9 +325,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
resource.samples != renderbuffer->GetSamples();
if (!needsCreate) {
resource.renderbuffer = renderbuffer;
// A new renderbuffer at a recycled address may adopt a compatible entry that
// was already stamped dead; it is alive again, so cancel the aging.
resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
return &resource;
}
// Respecify: park the old backing for aged destruction instead of destroying
// inline - it may still be referenced by in-flight command buffers.
DeferRenderbufferBackingRelease(resource);
resource.Destroy(m_device, m_allocator);
resource.renderbuffer = renderbuffer;
@@ -1178,6 +1239,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
void VkRenderPassManager::OnPresent() {
++m_frameCounter;
// Runs every frame boundary, ahead of the render-pass sweep gate below: the walk
// is O(#renderbuffer resources) — single digits in practice — and per-frame
// invocation keeps dead-resource reclaim latency at the aging bound instead of
// coupling it to renderbuffer *use* (the GetOrCreateRenderbufferResource call
// site never runs again once an app stops using renderbuffers).
CollectRenderbufferGarbage();
CollectDeferredRenderbufferReleases(/*destroyAll=*/false);
// Sweep occasionally; evict entries whose last use is far past every
// in-flight frame so their VkRenderPass/VkFramebuffer can be destroyed
// safely (RenderPassEntry's destructor releases the handles).
@@ -1187,6 +1256,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return;
}
// Collect the dying handles and notify once after the loop: pipelines hashed
// on them share the entries' >kRetireAgeFrames idleness (they are only bound
// by draws that hit those entries), so the observer may destroy them
// immediately - and a single batched notification costs one pipeline-cache
// scan instead of one per evicted pass.
Vector<VkRenderPass> destroyedRenderPasses;
const Uint64 activeHash = s_hasActiveRenderPass ? s_activeRenderPass.hash : 0;
for (auto it = m_renderPasses.begin(); it != m_renderPasses.end();) {
const Bool isActive = s_hasActiveRenderPass && it->first == activeHash;
@@ -1194,11 +1269,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
if (m_rpFastValid && m_rpFastRenderPassHash == it->first) {
m_rpFastValid = false;
}
destroyedRenderPasses.push_back(it->second.renderPass);
it = m_renderPasses.erase(it);
} else {
++it;
}
}
if (!destroyedRenderPasses.empty() && m_evictionObserver != nullptr) {
m_evictionObserver->OnRenderPassesDestroyed(destroyedRenderPasses);
}
}
Bool VkRenderPassManager::BeginRenderPass(VkCommandBuffer commandBuffer, RenderPassEntry& renderPassEntry) {
@@ -157,11 +157,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
class VkRenderPassManager {
public:
using HashType = Uint64;
// Notified once per OnPresent sweep with every aged-out entry's VkRenderPass
// value: pipelines are hashed on the raw handle, and once destroyed the value
// may be recycled for an incompatible pass, so dependent caches must purge
// everything keyed on them before any new pass can be created (the sweep and
// the notification run back-to-back with no creation in between; observers
// compare the values, never dereference them). Batched so a mass-idle cohort
// (shader-pack switch, dimension exit) costs the observer one pipeline-cache
// scan, not one per dying pass. The wholesale paths
// (Shutdown/RecreateSwapchain) do not notify - their callers already drop
// every pipeline outright.
class IEvictionObserver {
public:
virtual ~IEvictionObserver() = default;
virtual void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) = 0;
};
VkRenderPassManager(VkDevice device,
VkPhysicalDevice physicalDevice, VmaAllocator allocator, const VulkanRendererConfig& config,
VkClearManager& clearManager, VkTextureManager& textureManager, SwapchainObject& swapchainObject);
~VkRenderPassManager();
// Observer may be null (no notifications). Not owned.
void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
Bool Initialize();
void Shutdown();
@@ -192,6 +212,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
UnorderedMap<Uint64, RenderPassEntry> m_renderPasses;
// Monotonic frame counter (bumped in OnPresent) for render-pass cache aging.
Uint64 m_frameCounter = 0;
IEvictionObserver* m_evictionObserver = nullptr;
// Bumped whenever a renderbuffer VkImage is (re)created; together with the texture
// manager's image epoch this invalidates the render-pass fast path on any attachment
@@ -213,6 +234,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
public:
struct RenderbufferResource {
// deadSinceFrame sentinel: the owning weak reference has not been observed
// expired. Dead resources age past every in-flight frame before Destroy
// (see CollectRenderbufferGarbage); the GPU may still reference the image
// for frames-in-flight frames after the GL object dies.
static constexpr Uint64 kNeverObservedDead = UINT64_MAX;
WeakPtr<MG_State::GLState::RenderbufferObject> renderbuffer;
VkImage image = VK_NULL_HANDLE;
VmaAllocation allocation = nullptr;
@@ -224,6 +251,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
TextureInternalFormat internalFormat = TextureInternalFormat::Unknown;
Int samples = 0;
// m_frameCounter value at which the weak reference was first seen expired.
Uint64 deadSinceFrame = kNeverObservedDead;
void Destroy(VkDevice device, VmaAllocator allocator);
};
@@ -241,12 +270,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
ClearAttachmentPayload payload{};
};
// A superseded renderbuffer backing (glRenderbufferStorage respecify) parked
// until enough frame boundaries have passed that no in-flight command buffer
// can still reference it; destroyed in OnPresent (see RetireAgeFrames).
struct DeferredRenderbufferRelease {
VkImage image = VK_NULL_HANDLE;
VmaAllocation allocation = nullptr;
VkImageView view = VK_NULL_HANDLE;
Uint64 deferredAtFrame = 0;
};
UnorderedMap<MG_State::GLState::RenderbufferObject*, RenderbufferResource> m_renderbufferResources;
UnorderedMap<MG_State::GLState::RenderbufferObject*, PendingRenderbufferClear> m_pendingRenderbufferClears;
Vector<DeferredRenderbufferRelease> m_deferredRenderbufferReleases;
Bool HasPendingRenderbufferClear(
const MG_State::GLState::FramebufferAttachmentObject& attachment) const;
void CollectRenderbufferGarbage();
// Frame-boundary margin after which a resource last referenced by a retired
// GL object (or superseded backing) is provably past every in-flight frame.
Uint64 RetireAgeFrames() const;
void DeferRenderbufferBackingRelease(RenderbufferResource& resource);
void CollectDeferredRenderbufferReleases(Bool destroyAll);
static inline XXH64_state_t* m_hashState = XXH64_createState();
static inline ActiveRenderPassInfo s_activeRenderPass{};
@@ -89,6 +89,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
m_device = VK_NULL_HANDLE;
m_config = nullptr;
m_frameBoundaryCounter = 0;
}
void VkSamplerManager::OnFrameBoundary() {
++m_frameBoundaryCounter;
// Sweep occasionally; destroy samplers whose last use is far past every
// in-flight frame. Destroy and erase must stay atomic, or Shutdown would
// double-free the handle; an evicted key that recurs simply re-creates
// its sampler on the next miss.
constexpr Uint64 kSweepInterval = 256;
constexpr Uint64 kRetireAgeBoundaries = 1024;
if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
return;
}
for (auto it = m_samplers.begin(); it != m_samplers.end();) {
auto& entry = it->second;
if (m_frameBoundaryCounter - entry.lastUsedFrameBoundary > kRetireAgeBoundaries) {
if (m_device != VK_NULL_HANDLE && entry.handle != VK_NULL_HANDLE) {
vkDestroySampler(m_device, entry.handle, nullptr);
}
it = m_samplers.erase(it);
} else {
++it;
}
}
}
Uint64 VkSamplerManager::BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
@@ -137,6 +164,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering);
auto it = m_samplers.find(key);
if (it != m_samplers.end()) {
it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
return it->second.handle;
}
@@ -169,6 +197,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
entry.handle = vkSampler;
entry.externalIndex = sampler.GetExternalIndex();
entry.version = sampler.GetVersion();
entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
m_samplers[key] = entry;
return vkSampler;
}
@@ -36,12 +36,27 @@ public:
VkSampler GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
const MG_State::GLState::ITextureObject& texture,
Bool forceNearestFiltering = false);
// Frame boundary hook: ages the sampler cache and destroys samplers not used
// for many frames. The key hashes continuous float state (lodBias, LOD clamps,
// anisotropy), so an app animating those would otherwise mint an unbounded
// stream of never-destroyed VkSamplers and eventually exhaust the device's
// maxSamplerAllocationCount. A sampler idle for over a thousand frame
// boundaries cannot be referenced by any in-flight command buffer (frames in
// flight are single digits), and every descriptor set the GPU consumes is
// written that same frame with live handles (the per-binding resolve memo and
// descriptor-set reuse are both frame-reset), so destruction here needs no
// fence wait. Self-gated: one counter bump and compare except on sweep
// boundaries.
void OnFrameBoundary();
private:
struct SamplerCacheEntry {
VkSampler handle = VK_NULL_HANDLE;
Uint externalIndex = 0;
Uint16 version = 0;
// Frame boundary of the last cache hit; entries idle past the
// OnFrameBoundary retirement age have their VkSampler destroyed.
Uint64 lastUsedFrameBoundary = 0;
};
Uint64 BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
@@ -67,6 +82,8 @@ private:
Bool m_samplerAnisotropySupported = false;
Float m_maxSamplerAnisotropy = 1.0f;
UnorderedMap<Uint64, SamplerCacheEntry> m_samplers;
// Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
Uint64 m_frameBoundaryCounter = 0;
static inline XXH64_state_t* m_hashState = XXH64_createState();
};
} // namespace MobileGL::MG_Backend::DirectVulkan
@@ -627,6 +627,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
frameIndex, m_deferredViewReleases.size());
m_currentFrameIndex = frameIndex;
CollectDeferredReleases(frameIndex);
// Frame-boundary GC: every 64 frame boundaries (~1 s at 60 fps) bounds the reclaim
// latency for dead textures regardless of draw traffic — workloads that churn
// textures through clears/readbacks alone never reach the draw-gated
// CollectGarbage. Must run after CollectDeferredReleases above: the prune defers
// its releases into this frame's slot, which was just drained, so they are
// destroyed only after the slot's fence has been waited again one full frame-ring
// cycle from now (never while an in-flight frame may still reference them).
constexpr Uint32 kGcFrameInterval = 64;
++m_gcFrameCounter;
if (m_gcFrameCounter % kGcFrameInterval == 0) {
PruneDeadTextures();
}
}
void VkTextureManager::CollectAllDeferredReleases() {
const SizeT frameCount = std::min(m_deferredReleases.size(), m_deferredViewReleases.size());
for (SizeT frameIndex = 0; frameIndex < frameCount; ++frameIndex) {
CollectDeferredReleases(static_cast<Uint32>(frameIndex));
}
}
void VkTextureManager::EraseTrackedTexture(const TextureIdentity& identity) {
@@ -702,9 +722,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// construction introduces a new identity. Doing this unconditionally made every
// sampled-texture sync scan the entire alive-texture map per draw.
if (aliveIt == m_aliveObjects.end()) {
WeakPtr<MG_State::GLState::ITextureObject> aliveTexture;
const auto& liveTexture = MG_State::pGLContext->GetTextureObject(texture.GetExternalIndex());
if (liveTexture && liveTexture.get() == &texture) {
m_aliveObjects[identity] = WeakPtr<MG_State::GLState::ITextureObject>(liveTexture);
aliveTexture = liveTexture;
} else {
// The name lookup legally fails while the object is alive: the name was
// deleted with the texture still attached to an FBO (the attachment's
// SharedPtr keeps it alive), or the name was reused by a new texture, or
// this is a default texture object (name 0 lives outside the name map).
// Register through the object's own control block so the resource created
// below still participates in weak-expiry GC instead of becoming an
// orphan no reclamation path can reach until Shutdown.
aliveTexture = texture.weak_from_this();
}
if (!aliveTexture.expired()) {
m_aliveObjects[identity] = Move(aliveTexture);
PruneStaleTextureAliases(&texture);
}
}
@@ -1212,10 +1245,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
SizeT VkTextureManager::CollectGarbage() {
// Draw-gated stagger (1 in 256 calls): keeps the per-draw cost at one counter
// bump. The guaranteed reclaim path is the frame-boundary prune in BeginFrame;
// this remains as a cheap assist so draw-heavy workloads reclaim sooner.
m_gcCounter++;
if (m_gcCounter != 0) {
return 0;
}
return PruneDeadTextures();
}
SizeT VkTextureManager::PruneDeadTextures() {
// Erasing entries would dangle the raw TextureResource pointers memoized for the
// current draw; every call path (BeginFrame, and CollectGarbage at the top of a
// freshly opened draw-sync scope) runs before any memo entry is recorded.
MOBILEGL_ASSERT(m_drawSyncedThisDraw.empty(),
"PruneDeadTextures: draw-sync memo holds raw resource pointers an erase would dangle");
Vector<MG_State::GLState::ITextureObject*> expiredTextures;
expiredTextures.reserve(m_aliveObjects.size());
@@ -1227,7 +1272,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
for (auto* texture : expiredTextures) {
PruneStaleTextureAliases(texture);
}
return expiredTextures.size();
SizeT prunedCount = expiredTextures.size();
// Orphan sweep: after the pass above, m_aliveObjects holds only live entries.
// Registration in SyncTextureAndGetDescriptor cannot fail for a SharedPtr-owned
// texture (weak_from_this fallback), so a resource whose identity has no alive
// entry has no trackable owner: its GL-side object is gone, or was never
// shared-owned, in which case recreation on a later sync is the safe fallback.
// Destruction goes through the per-frame deferred queues, never immediate.
Vector<TextureIdentity> orphanIdentities;
for (auto it = m_textureResources.begin(); it != m_textureResources.end(); ++it) {
if (m_aliveObjects.find(it->first) == m_aliveObjects.end()) {
orphanIdentities.emplace_back(it->first);
}
}
for (const auto& identity : orphanIdentities) {
EraseTrackedTexture(identity);
}
prunedCount += orphanIdentities.size();
return prunedCount;
}
Bool VkTextureManager::SyncTexture(MG_State::GLState::ITextureObject &texture,
@@ -267,6 +267,10 @@ public:
Bool Initialize(const InitInfo& initInfo);
void Shutdown();
void BeginFrame(Uint32 frameIndex);
// Drains every frame slot's deferred image/view releases. Only valid when
// the caller has proven every queue submission complete; used by the
// present-less frame-boundary drain.
void CollectAllDeferredReleases();
TextureResource* SyncTextureAndGetDescriptor(
MG_State::GLState::ITextureObject& texture);
@@ -364,6 +368,7 @@ private:
static TextureIdentity MakeTextureIdentity(MG_State::GLState::ITextureObject* texture);
void EraseTrackedTexture(const TextureIdentity& identity);
void PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture);
SizeT PruneDeadTextures();
VkDevice m_device = VK_NULL_HANDLE;
VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
@@ -373,6 +378,9 @@ private:
Uint32 m_currentFrameIndex = 0;
Uint8 m_gcCounter = 0;
// Frame-boundary GC gate: counts BeginFrame calls, not draws, so texture churn
// through non-draw paths (FBO clears, readbacks) still reaches the prune.
Uint32 m_gcFrameCounter = 0;
// Active only between BeginDrawSyncScope/EndDrawSyncScope; identities of
// textures already fully synced in the current draw (small N -> flat scan).
Bool m_drawSyncScopeActive = false;
@@ -2529,6 +2529,11 @@ void main() {
m_shaderDrawParametersFeatureEnabled,
m_unformattedFloatStorageImagesEnabled);
MOBILEGL_ASSERT(m_programFactory != nullptr, "ProgramFactory creation failed.");
// Aging evictions (render passes and program entries) must purge the dependent
// pipeline / compute-pipeline / descriptor-set caches in the same step; both
// sweeps only run from the frame-boundary seams, long after initialization.
m_renderPassManager->SetEvictionObserver(this);
m_programFactory->SetEvictionObserver(this);
m_samplerManager = MakeUnique<VkSamplerManager>();
MOBILEGL_ASSERT(m_samplerManager != nullptr, "VkSamplerManager creation failed.");
@@ -2589,6 +2594,15 @@ void main() {
DestroyDeferredDepthMipmapCleanup();
DestroyComputePipelines();
// No sweep runs during teardown, but the observers point at this renderer
// and the factories die at different times below; disconnect them first.
if (m_renderPassManager) {
m_renderPassManager->SetEvictionObserver(nullptr);
}
if (m_programFactory) {
m_programFactory->SetEvictionObserver(nullptr);
}
m_pipelineFactory.reset();
ShutdownBlitResources();
ShutdownDepthMipmapResources();
@@ -6204,12 +6218,12 @@ void main() {
frame.hasCommandBufferRecorded = false;
frame.isCommandRecording = false;
// The wait proved every descriptor set this slot has in flight idle;
// rewind the reuse cursors so present-less readback loops stay
// bounded (Present is the only other rewind point).
if (m_uniformManager) {
m_uniformManager->BeginFrame(m_frameContext.GetCurrentFrameIndex());
}
// The wait proved every submission complete, so the full frame-boundary
// drain applies: descriptor cursors, transient arenas, deferred
// texture/buffer releases, retired command buffers and the converted
// vertex-stream cache all rewind here, keeping present-less readback
// loops bounded (Present is the only other drain point).
TryDrainFrameTransients();
return true;
}
@@ -7123,6 +7137,9 @@ void main() {
}
m_bufferManager.NotifyDeviceIdle();
OnSubmitsCompletedUpTo(m_submitCounter);
// The queue was just drained; take the free frame-boundary drain when
// nothing is recorded (present-less timer-query loops). No-op otherwise.
TryDrainFrameTransients();
return true;
}
@@ -7191,6 +7208,85 @@ void main() {
vkDestroyFence(m_device, record.fence, nullptr);
}
}
// Mid-frame-flushed command buffers whose submission just completed can
// be freed now; present-less flush loops have no other reclaim point.
m_frameContext.FreeRetiredCommandBuffersCompletedUpTo(m_completedSubmitCounter);
}
Bool VulkanRenderer::TryDrainFrameTransients() {
if (m_device == VK_NULL_HANDLE || m_frameContext.GetFrameCount() == 0) {
return false;
}
if (m_completedSubmitCounter != m_submitCounter) {
RefreshCompletedSubmits();
if (m_completedSubmitCounter != m_submitCounter) {
return false;
}
}
if (HasPendingRecordedWork()) {
return false;
}
// Every submission is complete and nothing recorded references the
// per-frame transients. Pure-reclaim work runs on every drain: it only
// releases memory that is provably dead, never invalidates anything a
// later draw would have to rebuild. Raise the buffer manager's
// completed floor first so busy-tracking reflects the proven idleness.
m_bufferManager.NotifyDeviceIdle();
const Uint32 frameIndex = m_frameContext.GetCurrentFrameIndex();
m_frameContext.FreeAllRetiredCommandBuffers();
for (Uint32 slot = 0; slot < m_deferredDepthMipmapCleanup.size(); ++slot) {
CollectDeferredDepthMipmapCleanup(slot);
}
if (m_textureManager) {
m_textureManager->CollectAllDeferredReleases();
}
m_bufferManager.CollectAllDeferredReleases();
// Descriptor cursors rewind on every drain (the pre-drain readback path
// already did exactly this), keeping fence/readback loops' set usage bounded.
if (m_uniformManager) {
m_uniformManager->BeginFrame(frameIndex);
}
// Frame-boundary-equivalent work - transient arena rewind (which invalidates
// the conversion cache) and the cache-aging clocks - is gated to every 8th
// drain since the last Present: a presenting app's mid-frame readbacks/waits
// must neither force re-conversion/re-upload churn for the rest of the frame
// nor multiply the aging rate (which would shrink the 1024-boundary retire
// window and thrash periodically-used pipelines/programs), while present-less
// loops still rewind the arena and age their caches every 8 iterations -
// bounded by 8 iterations' transient usage.
++m_drainsSinceLastPresent;
if ((m_drainsSinceLastPresent % 8) != 0) {
return true;
}
if (m_textureManager) {
m_textureManager->BeginFrame(frameIndex);
}
m_bufferManager.BeginFrame(frameIndex);
// The cached conversion slices point into the transient arena the
// BeginFrame above just rewound; drop them together.
m_convertedVertexStreams.clear();
if (m_renderPassManager) {
m_renderPassManager->OnPresent();
}
// The pipeline memo can survive across these boundaries (no per-frame reset
// on this path), so it must drop whenever the sweep destroys anything.
if (m_programFactory) {
m_programFactory->OnFrameBoundary();
}
if (m_pipelineFactory && m_pipelineFactory->OnFrameBoundary() > 0) {
m_lastPipelineValid = false;
m_lastPipelineResult = VK_NULL_HANDLE;
}
if (m_vertexInputStateFactory) {
m_vertexInputStateFactory->OnFrameBoundary();
}
if (m_samplerManager) {
m_samplerManager->OnFrameBoundary();
}
return true;
}
VkFence VulkanRenderer::AcquirePooledSubmitFence() {
@@ -7253,6 +7349,10 @@ void main() {
if (m_device == VK_NULL_HANDLE || m_graphicsQueue == VK_NULL_HANDLE || m_frameContext.GetFrameCount() == 0) {
return false;
}
// Non-blocking completion poll: gives flush-only workloads (no sync
// objects, no present) a point where finished submissions retire their
// pooled fences and mid-frame command buffers.
RefreshCompletedSubmits();
auto& frame = m_frameContext.GetCurrent();
if (!frame.isCommandRecording && !frame.hasCommandBufferRecorded) {
return false;
@@ -7277,6 +7377,14 @@ void main() {
return false;
}
// Command-buffer boundary: the pipeline memo must not survive it, or a
// pipeline bound only through memo hits is never re-stamped in the factory
// cache and the aging sweep could destroy it while the flushed submission
// still references it. Mirrors the drops at the readback and Present
// boundaries; costs one full pipeline lookup on the next draw.
m_lastPipelineValid = false;
m_lastPipelineResult = VK_NULL_HANDLE;
// The submitted command buffer may still be executing; recording must
// restart on a fresh one. If none can be allocated, fall back to
// draining this submission so reusing the buffer stays legal.
@@ -7328,6 +7436,10 @@ void main() {
const VkResult result = vkWaitForFences(m_device, 1, &record.fence, VK_TRUE, timeoutNs);
if (result == VK_SUCCESS) {
OnSubmitsCompletedUpTo(record.submitIndex);
// The wait already stalled the pipeline; if it happens to
// have drained everything (present-less fence loops), take
// the free frame-boundary drain. No-op otherwise.
TryDrainFrameTransients();
return true;
}
if (result != VK_TIMEOUT) {
@@ -7417,6 +7529,13 @@ void main() {
suspendedFrame.isCommandRecording = false;
suspendedFrame.hasCommandBufferRecorded = false;
m_lastPipelineValid = false;
// The dropped recording is never submitted, so once the fence
// poll shows the pre-suspension submissions complete the frame
// transients (descriptor sets, transient arenas, deferred
// releases, conversion caches) can rewind; without this a
// minimized-window app accumulates them for the whole
// suspension.
TryDrainFrameTransients();
MGLOG_D("Present skipped: no usable swapchain (zero-area window)");
return;
}
@@ -7430,6 +7549,22 @@ void main() {
MOBILEGL_ASSERT(m_imageIndexAcquired < m_swapchainObject.GetImageCount(),
"Present, acquired image index out of range");
m_renderPassManager->OnPresent();
// A real presented frame is the canonical aging cadence; mid-frame drains
// count against this and only age when presents stop coming.
m_drainsSinceLastPresent = 0;
// Age the content-addressed caches on the same frame-boundary cadence. Each
// keeps its own internal 256-sweep gate, so the per-frame cost is one counter
// increment and compare per cache; entries used by this frame's still-
// unsubmitted recording were stamped this boundary (every command-buffer
// boundary drops the pipeline memo, so the first draw of each recording
// performs a real, stamping lookup) and can never age out.
m_programFactory->OnFrameBoundary();
if (m_pipelineFactory->OnFrameBoundary() > 0) {
m_lastPipelineValid = false; // an aged-out pipeline may still be memoized
m_lastPipelineResult = VK_NULL_HANDLE;
}
m_vertexInputStateFactory->OnFrameBoundary();
m_samplerManager->OnFrameBoundary();
auto& frame = m_frameContext.GetCurrent();
auto* activeRenderPass = VkRenderPassManager::GetActiveRenderPass();
if (activeRenderPass)
@@ -8462,6 +8597,43 @@ void main() {
m_computePipelines.clear();
}
void VulkanRenderer::OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) {
if (m_pipelineFactory == nullptr) {
return;
}
// The render-pass sweep's >1024-boundary idle guarantee covers these pipelines
// too (they are only bound by draws that hit the dying entries), so the factory
// destroys them immediately. The memo must drop as well: it can hand out a
// cached handle without touching the factory.
if (m_pipelineFactory->EvictByRenderPasses(renderPasses) > 0) {
m_lastPipelineValid = false;
m_lastPipelineResult = VK_NULL_HANDLE;
}
}
void VulkanRenderer::OnProgramEvicted(ProgramFactory::HashType programHash,
VkDescriptorSetLayout descriptorSetLayout) {
// Same >1024-boundary idleness as the program entry: its compute pipeline is
// only dispatched, and its graphics pipelines only bound, through paths that
// stamp the entry, so immediate destruction is GPU-safe. (The graphics memo
// never holds compute pipelines; it only needs invalidating for the factory
// eviction below.)
const auto computeIt = m_computePipelines.find(programHash);
if (computeIt != m_computePipelines.end()) {
if (computeIt->second != VK_NULL_HANDLE && m_device != VK_NULL_HANDLE) {
vkDestroyPipeline(m_device, computeIt->second, nullptr);
}
m_computePipelines.erase(computeIt);
}
if (m_pipelineFactory != nullptr && m_pipelineFactory->EvictByProgramHash(programHash) > 0) {
m_lastPipelineValid = false;
m_lastPipelineResult = VK_NULL_HANDLE;
}
if (m_uniformManager != nullptr) {
m_uniformManager->OnDescriptorSetLayoutDestroyed(descriptorSetLayout);
}
}
VkPipeline VulkanRenderer::GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj) {
const auto it = m_computePipelines.find(programObj.hash);
if (it != m_computePipelines.end()) {
@@ -114,7 +114,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
};
class VulkanRenderer : public IBufferCopyCommandProvider, public FrameContext::IRecordingObserver {
class VulkanRenderer : public IBufferCopyCommandProvider,
public FrameContext::IRecordingObserver,
public VkRenderPassManager::IEvictionObserver,
public ProgramFactory::IEvictionObserver {
public:
VulkanRenderer(NativeWindowType window, const VulkanRendererConfig& cfg = {});
~VulkanRenderer();
@@ -131,6 +134,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// recording, before any render pass.
void OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) override;
// VkRenderPassManager::IEvictionObserver: the render-pass aging sweep just
// destroyed these VkRenderPasses; evict every graphics pipeline hashed on a
// dying handle (they share its >1024-boundary idleness, so immediate
// destruction is safe) and drop the last-pipeline memo if any went.
void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) override;
// ProgramFactory::IEvictionObserver: an aged-out program entry was
// destroyed; evict its compute pipeline and graphics pipelines (same
// idleness guarantee - they are only bound through draws/dispatches that
// stamp the program entry) and purge the descriptor-set cache entries
// keyed by its now-recyclable VkDescriptorSetLayout handle.
void OnProgramEvicted(ProgramFactory::HashType programHash,
VkDescriptorSetLayout descriptorSetLayout) override;
Bool SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
const DrawCmdParam& drawParams,
const IndexBufferView* pIndexBufferView = nullptr);
@@ -345,11 +362,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
VkFence AcquirePooledSubmitFence();
void DestroySubmitFencePool();
Bool HasPendingRecordedWork() const;
// Frame-boundary housekeeping for paths that never reach Present's
// tail (present-less readback loops, suspended presentation, blocking
// sync waits): runs the same per-frame drains Present performs, but
// only when every queue submission has been observed complete AND no
// recorded-but-unsubmitted commands exist - i.e. when CPU-GPU overlap
// is provably already zero. Never blocks (non-blocking fence poll
// only), so the presenting path's frames-in-flight pipelining is
// untouched. Returns true when the drain ran.
Bool TryDrainFrameTransients();
Vector<SubmitRecord> m_inFlightSubmits;
Vector<VkFence> m_freeSubmitFences;
Uint64 m_submitCounter = 0;
Uint64 m_completedSubmitCounter = 0;
// Drains since the last Present, gating the drain's frame-boundary-equivalent
// work (arena rewind + cache aging): a presenting app's mid-frame
// readbacks/waits must neither churn the transient caches nor accelerate the
// aging clocks, while present-less loops still cross a boundary every few
// iterations. Reset in Present.
Uint32 m_drainsSinceLastPresent = 0;
NativeWindowType m_window = 0;
void* m_platformDisplay = nullptr;
+29
View File
@@ -133,4 +133,33 @@ namespace MobileGL::MG_Impl::GLImpl {
values[0] = value;
}
}
void DestroyAllSyncObjects() {
// Detach the registry under the lock, release outside it. Entries the app
// already deleted were erased by DeleteSync, so nothing here double-frees;
// a DeleteSync racing this sweep finds an empty registry and returns. A
// thread still blocked inside ClientWaitSync/GetSynciv during teardown
// holds a raw SyncObject* these deletes invalidate - the same undefined
// race an app-driven DeleteSync already has.
UnorderedMap<GLsync, SyncObject*> orphans;
{
const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
orphans.swap(g_liveSyncObjects);
}
if (orphans.empty()) {
return;
}
// Both backends' DeleteSync only free the heap wrapper once their GL
// context/renderer is gone (generation/current-thread guards), so this is
// safe after the backend has released its EGL resources - but not after
// the function table itself is cleared.
const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
for (const auto& [_, syncObject] : orphans) {
if (backendDeleteSync && syncObject->backendHandle) {
backendDeleteSync(syncObject->backendHandle);
}
delete syncObject;
}
MGLOG_D("DestroyAllSyncObjects: reclaimed %zu sync object(s) the app left undeleted", orphans.size());
}
} // namespace MobileGL::MG_Impl::GLImpl
+8
View File
@@ -16,4 +16,12 @@ namespace MobileGL::MG_Impl::GLImpl {
void WaitSync(GLsync sync, GLbitfield flags, GLuint64 timeout);
void DeleteSync(GLsync sync);
void GetSynciv(GLsync sync, GLenum pname, GLsizei bufSize, GLsizei* length, GLint* values);
// Destroys every still-registered sync object exactly as DeleteSync would.
// GL requires syncs to die with their context; called only from full library
// teardown (DestroyImpl), where no context survives on any thread, so the
// process-global registry can be drained wholesale. Must run while the
// backend function table is still populated: each backend handle has to be
// released by the backend that created it, never by a later re-initialized
// one.
void DestroyAllSyncObjects();
} // namespace MobileGL::MG_Impl::GLImpl
@@ -15,7 +15,11 @@
#include <MG_Util/Math/VectorTypes.h>
namespace MobileGL::MG_State::GLState {
class ITextureObject {
// Texture objects are always SharedPtr-owned (TextureState creates every instance via
// MakeShared, including the per-target default objects). enable_shared_from_this lets
// backends that only receive a reference (e.g. syncing a name-deleted texture kept
// alive by an FBO attachment) still register a weak liveness reference for GC.
class ITextureObject : public std::enable_shared_from_this<ITextureObject> {
public:
using TargetEnum = TextureTarget;
virtual ~ITextureObject() = default;
+59
View File
@@ -33,6 +33,7 @@
#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
#include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
#include <MG_Util/Debug/Log.h>
#include <FastSTL/UnorderedMap.h>
namespace {
class DynamicParameterBackend final : public MobileGL::MG_Backend::BackendObject {
@@ -1736,3 +1737,61 @@ TEST(DirectGLESStateGuards, DefaultFramebufferBindGoesThroughShadow) {
FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 7); // must reach the driver again
EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 3u);
}
// FastSTL::unordered_map::erase(iterator) regression coverage. The open-addressing
// iterator constructor snaps forward from a tombstoned slot to the successor, so
// erase must NOT advance the rebuilt iterator again: the old double-advance skipped
// one live element per erase, and erasing the element in the highest occupied
// bucket pushed the returned index past bucket_count where it never compared equal
// to end() again - erase-while-iterating sweeps (pipeline/program cache eviction)
// then ran off the bucket array and fed garbage handles to vkDestroyPipeline
// (device crash on first mass eviction during world load).
TEST(FastSTLSanity, EraseWhileIteratingVisitsEveryElementExactlyOnce) {
FastSTL::unordered_map<MobileGL::Uint64, MobileGL::Uint64> map;
constexpr MobileGL::Uint64 kCount = 1000;
for (MobileGL::Uint64 key = 0; key < kCount; ++key) {
map.emplace(key * 0x9e3779b97f4a7c15ull, key);
}
ASSERT_EQ(map.size(), kCount);
MobileGL::SizeT visited = 0;
for (auto it = map.begin(); it != map.end();) {
it = map.erase(it);
++visited;
ASSERT_LE(visited, kCount); // old code: runaway past end / skipped entries
}
EXPECT_EQ(visited, kCount);
EXPECT_EQ(map.size(), 0u);
}
TEST(FastSTLSanity, EraseReturnsTheSuccessorElement) {
FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
for (MobileGL::Uint32 key = 1; key <= 64; ++key) {
map.emplace(key, key);
}
// Erasing every other visited element must still visit all 64 exactly once:
// the iterator returned by erase names the very next element, not one past it.
MobileGL::SizeT visited = 0;
MobileGL::SizeT erased = 0;
for (auto it = map.begin(); it != map.end();) {
++visited;
if ((visited & 1) != 0) {
it = map.erase(it);
++erased;
} else {
++it;
}
ASSERT_LE(visited, 64u);
}
EXPECT_EQ(visited, 64u);
EXPECT_EQ(map.size(), 64u - erased);
}
TEST(FastSTLSanity, ErasingTheOnlyElementReturnsEnd) {
FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
map.emplace(42u, 1u);
auto next = map.erase(map.begin());
EXPECT_EQ(next, map.end());
EXPECT_TRUE(map.empty());
}