[Merge] (P11): dev@c14283ee - scattered sub-image writes upload exactly their texels, Magma submits a frame's recording every 16384 draws with backpressure, and Magma grows descriptor pools geometrically and trims a quiet slot whole; the handle arm notes driver-side writes by handle and the wire arm splits recordings and keeps the trim's epoch peak across descriptor rewinds

This commit is contained in:
2026-09-29 12:11:37 -04:00
30 changed files with 2417 additions and 95 deletions
@@ -591,6 +591,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
m_samplerResolveMemoHighWater = 0;
m_setsPerFrame = setsPerFrame;
m_peakDescriptorSetsObserved = 0;
m_trimQuietEpochs = MG_Config::Features.MagmaDescriptorTrimFrames;
// The renderer records its first frame into slot 0 before any BeginFrame runs.
m_epochSlot = 0;
m_epochOpen = true;
m_descriptorSlotTrims = 0;
m_textureManager = textureManager;
m_samplerManager = samplerManager;
@@ -600,6 +605,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
frame.activeDescriptorPoolIndex = 0;
frame.allocatedSetsThisFrame = 0;
frame.peakAllocatedSetsThisFrame = 0;
frame.capacitySets = m_setsPerFrame;
frame.quietEpochs = 0;
frame.descriptorPools.clear();
VkDescriptorPool initialPool = VK_NULL_HANDLE;
@@ -654,6 +661,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
frame.activeDescriptorPoolIndex = 0;
frame.allocatedSetsThisFrame = 0;
frame.peakAllocatedSetsThisFrame = 0;
frame.capacitySets = 0;
frame.quietEpochs = 0;
}
m_frames.clear();
#if MOBILEGL_BUILD_DISAGGREGATED
@@ -662,6 +671,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
m_wirePrivateStoragePlaceholders = 0;
m_wireInvalidStorageImagesBindNull = false;
#endif
m_epochSlot = 0;
m_epochOpen = false;
m_bufferManager = nullptr;
m_programFactory = nullptr;
@@ -701,6 +712,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
"UniformDescriptorBinder: new descriptor set peak observed=%u (base setsPerFrame=%u, frame=%u, pools=%zu)",
m_peakDescriptorSetsObserved, m_setsPerFrame, frameIndex, frame.descriptorPools.size());
}
// The epoch that just ended is the one of the slot begun last - this slot again on the
// drain path, the previous frame's slot on the Present path. Its peak is still unreset.
if (m_epochOpen && m_epochSlot < m_frames.size()) {
NoteEndedEpochForTrim(m_frames[m_epochSlot].peakAllocatedSetsThisFrame);
}
m_epochSlot = frameIndex;
m_epochOpen = true;
// No pending command buffer references this slot's sets here (see TrimFrameDescriptorPools
// for why that holds at both of BeginFrame's call sites), so a quiet grown slot can be
// handed back whole.
if (m_trimQuietEpochs != 0 && frame.descriptorPools.size() > 1 &&
frame.capacitySets >= kDescriptorTrimMinCapacity && frame.quietEpochs >= m_trimQuietEpochs) {
TrimFrameDescriptorPools(frame, frameIndex);
}
frame.activeDescriptorPoolIndex = 0;
frame.allocatedSetsThisFrame = 0;
frame.peakAllocatedSetsThisFrame = 0;
@@ -709,11 +734,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
// The frame's descriptor sets are recycled above, so last frame's reuse targets
// are gone: start the per-draw descriptor-reuse cache fresh this frame.
for (auto& entry : m_descriptorReuseMemo) {
entry.valid = false;
}
m_fastRebindMemo.valid = false;
m_lastBindValid = false;
InvalidateDescriptorSetMemos();
// Re-fingerprint the bound sampler set fresh this frame so any GL object address
// reuse cannot outlive a single frame (see SamplerResolveMemo). Only the entries a
// resolve has actually written can be valid, so the high-water mark bounds the
@@ -747,10 +768,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
entry.second.cursor = 0;
}
frame.allocatedSetsThisFrame = 0;
frame.peakAllocatedSetsThisFrame = 0;
for (auto& entry : m_descriptorReuseMemo) entry.valid = false;
m_fastRebindMemo.valid = false;
m_lastBindValid = false;
// The PEAK is not reset: a rewind ends a window of the epoch, not the epoch. The trim scores
// an epoch by the most sets the slot held at once in it (NoteEndedEpochForTrim at the next
// BeginFrame), and on this arm that is the largest window - resetting here would score a
// burst frame by its last window alone and trim a slot that re-grows every frame.
InvalidateDescriptorSetMemos();
const Uint32 touchedBindings =
std::min<Uint32>(m_samplerResolveMemoHighWater, static_cast<Uint32>(m_samplerResolveMemo.size()));
for (Uint32 binding = 0; binding < touchedBindings; ++binding) {
@@ -761,6 +783,68 @@ namespace MobileGL::MG_Backend::DirectVulkan {
}
#endif
void UniformManager::InvalidateDescriptorSetMemos() {
// Every VkDescriptorSet handle held outside the per-layout caches - in this backend they all
// live in this class. A trim can hand a destroyed set's handle value to a fresh allocation,
// so the bind-dedup shadow must drop too, not only the two reuse memos.
for (auto& entry : m_descriptorReuseMemo) {
entry.valid = false;
}
m_fastRebindMemo.valid = false;
m_lastBindValid = false;
}
void UniformManager::NoteEndedEpochForTrim(Uint32 epochPeakSets) {
for (auto& slot : m_frames) {
if (static_cast<Uint64>(epochPeakSets) * kDescriptorTrimHeadroom > slot.capacitySets) {
slot.quietEpochs = 0;
} else if (slot.quietEpochs < std::numeric_limits<Uint32>::max()) {
++slot.quietEpochs;
}
}
}
void UniformManager::TrimFrameDescriptorPools(FrameResources& frame, Uint32 frameIndex) {
const SizeT poolsBefore = frame.descriptorPools.size();
const Uint64 capacityBefore = frame.capacitySets;
// Holders first: the per-layout cache and every memo that may name one of these sets.
frame.descriptorSetCacheByLayout.clear();
InvalidateDescriptorSetMemos();
for (SizeT i = 1; i < frame.descriptorPools.size(); ++i) {
if (frame.descriptorPools[i].handle != VK_NULL_HANDLE) {
vkDestroyDescriptorPool(m_device, frame.descriptorPools[i].handle, nullptr);
}
}
frame.descriptorPools.resize(1);
auto& base = frame.descriptorPools[0];
// Frees every set allocated from it; the spec defines no failure for vkResetDescriptorPool.
vkResetDescriptorPool(m_device, base.handle, 0);
base.allocatedSets = 0;
frame.activeDescriptorPoolIndex = 0;
frame.capacitySets = base.maxSets;
frame.quietEpochs = 0;
++m_descriptorSlotTrims;
MGLOG_D("UniformDescriptorBinder: frame %u trimmed its descriptor pools (%zu pools / %llu sets -> 1 / %u)",
frameIndex, poolsBefore, static_cast<unsigned long long>(capacityBefore), base.maxSets);
}
DescriptorPoolCensus UniformManager::GetDescriptorPoolCensus() const {
DescriptorPoolCensus census;
census.available = true;
census.frameSlots = static_cast<uint32_t>(m_frames.size());
for (const auto& frame : m_frames) {
const auto poolCount = static_cast<uint32_t>(frame.descriptorPools.size());
census.pools += poolCount;
census.maxPoolsInOneSlot = std::max(census.maxPoolsInOneSlot, poolCount);
census.capacitySets += frame.capacitySets;
for (const auto& cacheEntryPair : frame.descriptorSetCacheByLayout) {
census.cachedSets += cacheEntryPair.second.sets.size();
}
}
census.slotTrims = m_descriptorSlotTrims;
return census;
}
void UniformManager::OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout) {
SizeT purgedSets = 0;
for (auto& frame : m_frames) {
@@ -2831,32 +2915,49 @@ namespace MobileGL::MG_Backend::DirectVulkan {
return true;
}
Uint32 UniformManager::NextDescriptorPoolMaxSets(const Vector<DescriptorPoolBucket>& pools, Bool updateAfterBind,
Uint32 initialMaxSets) {
const Uint32 minSets = std::max<Uint32>(1, initialMaxSets);
const Uint32 maxSets = std::max(minSets, kMaxDescriptorPoolSets);
Uint32 largest = 0;
for (const auto& pool : pools) {
if (pool.updateAfterBind == updateAfterBind) {
largest = std::max(largest, pool.maxSets);
}
}
if (largest == 0) {
return minSets;
}
const Uint64 doubled = static_cast<Uint64>(largest) * 2;
return static_cast<Uint32>(std::clamp<Uint64>(doubled, minSets, maxSets));
}
Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind) {
if (frame.descriptorPools.empty()) {
return false;
}
const auto matchingBucket = std::find_if(
frame.descriptorPools.begin(), frame.descriptorPools.end(),
[updateAfterBind](const DescriptorPoolBucket& candidate) { return candidate.updateAfterBind == updateAfterBind; });
const Uint32 currentMaxSets = matchingBucket != frame.descriptorPools.end()
? std::max<Uint32>(1, matchingBucket->maxSets)
: m_setsPerFrame;
const Uint32 grownMaxSets = currentMaxSets <= (std::numeric_limits<Uint32>::max() / 2) ? (currentMaxSets * 2)
: currentMaxSets;
Uint32 grownMaxSets = NextDescriptorPoolMaxSets(frame.descriptorPools, updateAfterBind, m_setsPerFrame);
VkDescriptorPool grownPool = VK_NULL_HANDLE;
if (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u -> %u sets)",
currentMaxSets, grownMaxSets);
return false;
// A large pool is a large driver allocation; if the device refuses one, settle for less
// rather than failing the draw - halving down to the base size, where the old policy lived.
while (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
if (grownMaxSets <= m_setsPerFrame) {
MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u sets)",
grownMaxSets);
return false;
}
grownMaxSets = std::max(m_setsPerFrame, grownMaxSets / 2);
}
frame.descriptorPools.push_back({grownPool, grownMaxSets, 0, updateAfterBind});
frame.activeDescriptorPoolIndex = static_cast<Uint32>(frame.descriptorPools.size() - 1);
MGLOG_D(
"UniformDescriptorBinder: frame %u descriptor pool exhausted, grew pool (%u -> %u sets), poolCount=%zu",
frameIndex, currentMaxSets, grownMaxSets, frame.descriptorPools.size());
frame.capacitySets += grownMaxSets;
// Growing is the opposite of quiet: a trim must wait a full quiet run from here.
frame.quietEpochs = 0;
MGLOG_D("UniformDescriptorBinder: frame %u descriptor pool exhausted, grew a %u-set pool (updateAfterBind=%d), "
"poolCount=%zu",
frameIndex, grownMaxSets, updateAfterBind ? 1 : 0, frame.descriptorPools.size());
return true;
}
@@ -14,6 +14,7 @@
#include "VkSamplerManager.h"
#include "VkTextureManager.h"
#include "WirePlaceholderKey.h"
#include "../DescriptorPoolCensus.h"
#include "../VkIncludes.h"
#include <Includes.h>
@@ -84,9 +85,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// 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.
// references its sets. This is the only PER-LAYOUT eviction path for the
// caches - a live layout's entry must never be purged on its own (its sets would be
// unreachable pool slots), so there is deliberately no age-based sweep here. The
// other path, TrimFrameDescriptorPools, drops a slot's whole cache together with
// the pools behind it.
void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
// One record per visited CombinedImageSampler DESCRIPTOR (post fallback substitution,
// in binding order, and within a binding in array-element order): the resolved texture
@@ -158,13 +161,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
static Bool ProgramSamplesOnlySingleLevelTextures(const MagmaProgramSource& program,
const ProgramFactory::VkProgramObject& programObj);
private:
// ---- per-frame descriptor pools ----------------------------------------------------
// Each frame slot owns a list of pools. The first is the base pool (non-update-after-bind,
// setsPerFrame sets) and is never destroyed while the manager lives; every later one was
// added by GrowFrameDescriptorPool when an allocation found no room. A pool serves one
// updateAfterBind flavour, because a set can only come from a pool created with the flag
// its layout was.
struct DescriptorPoolBucket {
VkDescriptorPool handle = VK_NULL_HANDLE;
Uint32 maxSets = 0;
Uint32 allocatedSets = 0;
Bool updateAfterBind = false;
};
// One pool's maxSets never exceeds this. It bounds one pool's driver memory: CreateDescriptorPool
// declares 8 descriptors of each of 6 types per set, so a 4096-set pool asks for 196,608
// descriptors - about 12 MiB at 64 bytes each on a driver that backs every declared one.
static constexpr Uint32 kMaxDescriptorPoolSets = 4096;
// The growth policy, pure so the host tests can drive it: the size of the next pool a slot
// adds for `updateAfterBind` layouts, given the slot's existing pools. Twice the LARGEST pool
// of that flavour, capped at kMaxDescriptorPoolSets; `initialMaxSets` when the slot has no
// pool of that flavour yet. So a frame that needs N sets adds O(log N) pools while N stays
// under the cap, and one more pool per kMaxDescriptorPoolSets sets beyond it. Taking the
// largest (not the first) pool is the point: the first same-flavour pool is the oldest,
// smallest one, and doubling IT made every grown pool the same size.
static Uint32 NextDescriptorPoolMaxSets(const Vector<DescriptorPoolBucket>& pools, Bool updateAfterBind,
Uint32 initialMaxSets);
// Trim policy (see TrimFrameDescriptorPools). A slot is "quiet" in an epoch whose peak set
// count times kDescriptorTrimHeadroom still fits its capacity; a slot that has been quiet for
// MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES consecutive epochs, and holds at least
// kDescriptorTrimMinCapacity sets, is trimmed. The headroom is what keeps the trim from
// thrashing: geometric growth leaves a re-grown slot at least a third full (the pool a growth
// adds is at most twice the largest before it, so at most twice everything the slot held, and
// the epoch that grew it needed more than that), so the epoch that grew a slot is never quiet
// for it. The floor keeps a slot sized for ordinary
// frames from being trimmed at all - the trim exists to give back a burst.
static constexpr Uint32 kDescriptorTrimHeadroom = 4;
static constexpr Uint32 kDescriptorTrimMinCapacity = 1024;
// For tests (DescriptorPoolCensus.h): pools, capacity, cached sets and trims over every slot.
DescriptorPoolCensus GetDescriptorPoolCensus() const;
private:
// A cached descriptor set together with the pool it was allocated from, so a
// layout-destroyed purge can vkFreeDescriptorSets it back and credit the
@@ -239,6 +275,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
void DestroyWirePlaceholderImage(WirePlaceholderImage& image) const;
#endif
// descriptorSetCacheByLayout is the per-frame set cache: BeginFrame rewinds every cursor, so
// a slot re-writes the sets it allocated in earlier frames instead of allocating new ones, and
// the cache only ever grows - except through TrimFrameDescriptorPools, which gives a slot's
// whole burst back at once (and OnDescriptorSetLayoutDestroyed, one dead layout at a time).
struct FrameResources {
Vector<DescriptorPoolBucket> descriptorPools;
UnorderedMap<VkDescriptorSetLayout, DescriptorSetCacheEntry> descriptorSetCacheByLayout;
@@ -252,6 +292,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint32 activeDescriptorPoolIndex = 0;
Uint32 allocatedSetsThisFrame = 0;
Uint32 peakAllocatedSetsThisFrame = 0;
// Sum of maxSets over descriptorPools, kept in step by Initialize, Grow and Trim.
Uint64 capacitySets = 0;
// Consecutive epochs (see NoteEndedEpochForTrim) that were quiet for THIS slot's capacity.
Uint32 quietEpochs = 0;
};
static Bool ResolveSamplerTexture(const MagmaProgramSource& program,
@@ -395,6 +439,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const Vector<Uint32>& dynamicOffsets);
Bool CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const;
Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind);
// An epoch is the stretch between two BeginFrame calls: one frame, or the part of one between
// two drains (VulkanRenderer::TryDrainFrameTransients). Called by BeginFrame with the epoch
// that just ended, it updates every slot's quiet count - a burst in any slot counts against
// all of them, so bursts that land in the other slots still hold a slot's trim off.
void NoteEndedEpochForTrim(Uint32 epochPeakSets);
// Gives a slot's whole descriptor burst back: clears its per-layout cache, destroys every
// grown pool and resets the base pool, so the slot re-grows geometrically from setsPerFrame.
// Whole-slot rather than per-set: vkFreeDescriptorSets fragments a pool and can never empty
// a grown one. Called only from BeginFrame, which runs only when no pending command buffer
// references the slot's sets:
// - Present path (VulkanRenderer::Present, after WaitAndAcquireNextImage): the slot's
// imageInFlightFence has been waited. That fence rides the slot's LAST submission - the
// Present submit - and every mid-frame submission of the slot (FlushPendingCommands,
// SplitOversizedRecording, readbacks) went to the same queue earlier, so the fence's
// signal covers them (vkQueueSubmit's fence scope includes all earlier submissions).
// - Drain path (VulkanRenderer::TryDrainFrameTransients): every submission has been
// observed complete and no recording is open.
// Sets are per slot: a slot's sets are only ever bound into recordings made while it is the
// current slot, so trimming one slot never touches another slot's in-flight work.
void TrimFrameDescriptorPools(FrameResources& frame, Uint32 frameIndex);
// Drops every handle this class holds to a descriptor set outside the per-layout caches.
void InvalidateDescriptorSetMemos();
VkResult AllocateDescriptorSetsFromActivePool(
Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet);
VkResult AcquireDescriptorSet(Uint32 frameIndex,
@@ -412,6 +478,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
Uint32 m_maxBindings = 0;
Uint32 m_setsPerFrame = 0;
Uint32 m_peakDescriptorSetsObserved = 0;
// MOBILEGL_MAGMA_DESCRIPTOR_TRIM_FRAMES, latched at Initialize (0 = never trim).
Uint32 m_trimQuietEpochs = 0;
// The slot whose epoch is open (the last one BeginFrame was called for); valid once
// m_epochOpen is set. Its peak is the ended epoch's peak at the next BeginFrame.
Uint32 m_epochSlot = 0;
Bool m_epochOpen = false;
Uint64 m_descriptorSlotTrims = 0;
VkTextureManager* m_textureManager = nullptr;
VkSamplerManager* m_samplerManager = nullptr;
mutable SharedPtr<MG_State::GLState::ITextureObject> m_fallbackTexture2D;
@@ -4121,11 +4121,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
IntVec3 regionLo = {0, 0, 0};
IntVec3 regionSize = {0, 0, 0};
SizeT texelBytes = 0;
// Scatter refinement of the single dirty box: when the storage's rect
// list reports the writes' true footprint (~100 sprites whose union box
// spans the whole atlas), each rect is staged tightly and copied with
// its own VkBufferImageCopy in ONE vkCmdCopyBufferToImage. Empty means
// "stage the one box above". Only set while subRegion.
// The writes' exact footprint when it is more than one box (~100 sprites
// whose union box spans the whole atlas): each rect is staged tightly and
// copied with its own VkBufferImageCopy in ONE vkCmdCopyBufferToImage.
// Empty means "stage the one box above". Only set while subRegion.
Vector<MG_State::GLState::MipmapDirtyRegion> rects;
};
@@ -4137,6 +4136,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
targets.push_back(uploadTarget);
}
const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(mipmapTexture.GetFormat());
Vector<MG_State::GLState::MipmapDirtyRegion> dirtyFootprint;
VkDeviceSize stagingSize = 0;
for (const TextureUploadTarget target : targets) {
@@ -4198,8 +4198,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const SizeT texelCount = static_cast<SizeT>(texelSize.x()) *
static_cast<SizeT>(texelSize.y()) *
static_cast<SizeT>(std::max(texelSize.z(), 1));
if (!region.Empty() && !region.CoversWholeLevel(texelSize) && texelCount > 0 &&
levelBytes % texelCount == 0) {
// Exactly the texels the client wrote, never their union box or the storage's
// rect list (which merges touching writes into their bounding box, folds when
// full and is withheld past 3/4 of the box): the image may hold what the GPU
// drew or cleared anywhere else, and the shadow is stale there. Each box is one
// more VkBufferImageCopy in the same vkCmdCopyBufferToImage, which costs next
// to nothing, so there is no cheaper shape worth being wrong for. A union box
// that spans the level with gaps in it is not a whole-level write either.
mipmapTexture.GetStorageDirtyFootprint(target, level, dirtyFootprint);
if (!region.Empty() && (dirtyFootprint.size() >= 2 || !region.CoversWholeLevel(texelSize)) &&
texelCount > 0 && levelBytes % texelCount == 0) {
uploadItem.subRegion = true;
uploadItem.regionLo = region.lo;
uploadItem.regionSize = {region.hi.x() - region.lo.x(), region.hi.y() - region.lo.y(),
@@ -4209,17 +4217,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
static_cast<SizeT>(uploadItem.regionSize.y()) *
static_cast<SizeT>(uploadItem.regionSize.z()) *
uploadItem.texelBytes;
// Scatter refinement: the storage only hands out its rect list
// when the rects' summed area is materially smaller than the
// union box (0 otherwise), so taking it always stages fewer
// bytes than the box - the very amplification this path exists
// to avoid paying twice.
MG_State::GLState::MipmapDirtyRegion
dirtyRects[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
const SizeT dirtyRectCount = mipmapTexture.GetStorageDirtyRects(
target, level, dirtyRects, MG_State::GLState::MipmapStorage::kMaxDirtyRects);
if (dirtyRectCount >= 2) {
uploadItem.rects.assign(dirtyRects, dirtyRects + dirtyRectCount);
// Several boxes: each is staged tightly and copied on its own. One box is
// the union box above.
if (dirtyFootprint.size() >= 2) {
uploadItem.rects.assign(dirtyFootprint.begin(), dirtyFootprint.end());
SizeT rectTexels = 0;
for (const auto& rect : uploadItem.rects) {
rectTexels += rect.TexelCount();
@@ -4479,6 +4480,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
const Bool depthSelectsArrayLayer = outResource.viewType == VK_IMAGE_VIEW_TYPE_1D_ARRAY ||
outResource.viewType == VK_IMAGE_VIEW_TYPE_2D_ARRAY ||
outResource.viewType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
Vector<VkBufferImageCopy> rectCopies;
for (const auto& item : uploadItems) {
if (!item.rects.empty()) {
// Multi-rect item: one VkBufferImageCopy per rect, all submitted in a
@@ -4486,7 +4488,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// construction, so no two copies write the same texels. Multi-rect
// implies subRegion, which implies a plain color aspect - the combined
// depth-stencil split below can never see one of these.
VkBufferImageCopy rectCopies[MG_State::GLState::MipmapStorage::kMaxDirtyRects];
rectCopies.resize(item.rects.size());
Uint32 rectCopyCount = 0;
VkDeviceSize runningOffset = item.offset;
for (const auto& rect : item.rects) {
@@ -4517,7 +4519,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
runningOffset += static_cast<VkDeviceSize>(rect.TexelCount() * item.texelBytes);
}
vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, rectCopyCount, rectCopies);
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, rectCopyCount, rectCopies.data());
continue;
}
const Uint32 depthOrLayers = item.texelSize.z() > 0 ? static_cast<Uint32>(item.texelSize.z()) : 1u;
@@ -7083,6 +7083,10 @@ void main() {
Bool VulkanRenderer::SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
const DrawCmdParam& drawParams,
const IndexBufferView* pIndexBufferView) {
// First, before anything of this draw is recorded or resolved - on the wire arm too: a
// split server records every draw its client sends into the same frame command buffer, so
// the same loading frame grows it the same way there.
SplitOversizedRecording();
#if MOBILEGL_BUILD_DISAGGREGATED
if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
RewindWireDescriptorSetsIfDue();
@@ -7718,6 +7722,8 @@ void main() {
}
void VulkanRenderer::DispatchCompute(GLuint numGroupsX, GLuint numGroupsY, GLuint numGroupsZ) {
// Before the wire branch, for SetupDraw's reason.
SplitOversizedRecording();
#if MOBILEGL_BUILD_DISAGGREGATED
if (MG_Config::Transport != MG_Config::TransportMode::Monolith) {
RewindWireDescriptorSetsIfDue();
@@ -7779,6 +7785,7 @@ void main() {
}
void VulkanRenderer::DispatchComputeIndirect(GLintptr indirect) {
SplitOversizedRecording();
m_textureManager->CollectGarbage();
auto& frame = m_frameContext.GetCurrent();
// See DispatchCompute: the dispatch accessor, not the draw one.
@@ -14198,6 +14205,54 @@ void main() {
return FlushPendingCommands();
}
void VulkanRenderer::SplitOversizedRecording() {
const Uint32 maxDraws = MG_Config::Features.MagmaMaxDrawsPerCommandBuffer;
if (maxDraws == 0) {
return;
}
// Counted before the check, so a buffer carries at most maxDraws calls before the one
// that splits it. OnFrameCommandRecordingBegan resets the count on every new recording.
if (++m_drawsInRecording <= maxDraws || !HasPendingRecordedWork()) {
return;
}
// Nothing of the current call is recorded yet, so the split falls between two GL calls,
// exactly where a glFlush would put it: the render pass ends with STORE, the next draw
// re-begins it with LOAD, and queries and transform feedback are bracketed per draw.
if (!FlushPendingCommands()) {
return;
}
// Fresh command buffer: the same reset every other mid-frame flush site applies.
m_lastSampledSetValid = false;
++m_oversizedRecordingSplits;
MGLOG_I_ONCE("DirectVulkan: a frame recorded more than %u draws into one command buffer; submitting it and "
"continuing on a fresh one (MOBILEGL_MAGMA_MAX_DRAWS_PER_COMMAND_BUFFER, 0 = unbounded)",
maxDraws);
MGLOG_D("DirectVulkan: split an oversized recording (%llu so far, submit index %llu)",
static_cast<unsigned long long>(m_oversizedRecordingSplits),
static_cast<unsigned long long>(m_submitCounter));
// Splitting only bounds the driver's command memory if the submitted buffers are freed
// as fast as they are produced. A GPU that falls behind would park them all on the
// retired list, which is the same growth spread over more buffers - so hold the queue
// to frames-in-flight + 1 outstanding submissions, the steady state of a pipelined
// frame plus this split. Wait on the fence directly rather than through
// WaitForSubmitIndex: its opportunistic TryDrainFrameTransients may rewind the transient
// arena, and this runs at the start of a draw whose caller may already hold slices of it.
const SizeT maxOutstanding = static_cast<SizeT>(m_frameContext.GetFrameCount()) + 1;
while (m_inFlightSubmits.size() > maxOutstanding) {
const SubmitRecord oldest = m_inFlightSubmits.front();
if (oldest.fence == VK_NULL_HANDLE) {
break;
}
const VkResult waitResult = vkWaitForFences(m_device, 1, &oldest.fence, VK_TRUE, UINT64_MAX);
if (waitResult != VK_SUCCESS) {
MGLOG_E_ONCE("SplitOversizedRecording: vkWaitForFences returned %d", waitResult);
break;
}
OnSubmitsCompletedUpTo(oldest.submitIndex);
}
}
Bool VulkanRenderer::WaitForSubmitIndex(Uint64 submitIndex, Uint64 timeoutNs, Bool flushIfPending) {
if (IsSubmitIndexComplete(submitIndex)) {
return true;
@@ -14243,6 +14298,7 @@ void main() {
}
void VulkanRenderer::OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) {
m_drawsInRecording = 0;
// Dynamic state does not survive a command-buffer boundary.
ResetDynamicStateShadow();
InvalidateSetupDrawSnapshots();
@@ -327,6 +327,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// signal as soon as the GPU actually finishes, which MC 1.21.5's
// fence-paced ring buffers rely on to recycle their space.
Uint64 GetSyncPointSubmitIndex() const;
// For tests: the per-frame descriptor pools (see DescriptorPoolCensus.h).
DescriptorPoolCensus GetDescriptorPoolCensus() const {
return m_uniformManager ? m_uniformManager->GetDescriptorPoolCensus() : DescriptorPoolCensus{};
}
// Non-blocking: polls outstanding submission fences and reports
// whether every submission up to `submitIndex` has completed.
Bool IsSubmitIndexComplete(Uint64 submitIndex);
@@ -736,6 +740,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// only), so the presenting path's frames-in-flight pipelining is
// untouched. Returns true when the drain ran.
Bool TryDrainFrameTransients();
// Bounded recording: called at the top of every GL draw and dispatch, before anything
// of that call is recorded. Once the open frame command buffer holds
// MagmaMaxDrawsPerCommandBuffer draws/dispatches, submits it (FlushPendingCommands) so
// the call continues on a fresh one, then waits until at most frames-in-flight + 1
// submissions are outstanding. A driver backs a command buffer with GPU memory it only
// returns when the buffer is freed (Adreno: one 16 KiB mapping per chunk), so a frame
// that is never split grows it without bound - rd12's loading frame records ~1M draws
// into one buffer and runs the process out of vm.max_map_count.
void SplitOversizedRecording();
Vector<SubmitRecord> m_inFlightSubmits;
Vector<VkFence> m_freeSubmitFences;
@@ -747,6 +760,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
// aging clocks, while present-less loops still cross a boundary every few
// iterations. Reset in Present.
Uint32 m_drainsSinceLastPresent = 0;
// GL draws/dispatches recorded into the open frame command buffer; reset whenever a
// recording begins (OnFrameCommandRecordingBegan). See SplitOversizedRecording.
Uint32 m_drawsInRecording = 0;
Uint64 m_oversizedRecordingSplits = 0;
NativeWindowType m_window = 0;
void* m_platformDisplay = nullptr;