Scarfski/filament
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Revert "Metal: implement more accurate buffer tracking (#7839)"

Browse Source 
This reverts commit 54a800a25d.
This commit is contained in:
Benjamin Doherty
2024-05-24 13:11:10 -07:00
parent d56f769d4d
commit 11ecaa2fbf
6 changed files with 77 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
filament/backend/src/metal/MetalBuffer.h
View File

@@ -65,12 +65,9 @@ private:
const char* mName;
};
#ifndef FILAMENT_METAL_BUFFER_TRACKING
#define FILAMENT_METAL_BUFFER_TRACKING 0
#endif
class MetalBufferTracking {
class TrackedMetalBuffer {
public:
static constexpr size_t EXCESS_BUFFER_COUNT = 30000;
enum class Type {
@@ -94,57 +91,66 @@ public:
}
}
#if FILAMENT_METAL_BUFFER_TRACKING
static void initialize() {
static dispatch_once_t onceToken;
dispatch_once(&onceToken, ^{
for (size_t i = 0; i < TypeCount; i++) {
aliveBuffers[i] = [NSHashTable weakObjectsHashTable];
}
});
}
static void setPlatform(MetalPlatform* p) { platform = p; }
static void track(id<MTLBuffer> buffer, Type type) {
TrackedMetalBuffer() noexcept : mBuffer(nil) {}
TrackedMetalBuffer(nullptr_t) noexcept : mBuffer(nil) {}
TrackedMetalBuffer(id<MTLBuffer> buffer, Type type) : mBuffer(buffer), mType(type) {
assert_invariant(type != Type::NONE);
if (UTILS_UNLIKELY(getAliveBuffers() >= EXCESS_BUFFER_COUNT)) {
if (platform && platform->hasDebugUpdateStatFunc()) {
platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
MetalBufferTracking::getAliveBuffers());
if (buffer) {
aliveBuffers[toIndex(type)]++;
mType = type;
if (getAliveBuffers() >= EXCESS_BUFFER_COUNT) {
if (platform && platform->hasDebugUpdateStatFunc()) {
platform->debugUpdateStat("filament.metal.excess_buffers_allocated",
TrackedMetalBuffer::getAliveBuffers());
}
}
}
[aliveBuffers[toIndex(type)] addObject:buffer];
}
~TrackedMetalBuffer() {
if (mBuffer) {
assert_invariant(mType != Type::NONE);
aliveBuffers[toIndex(mType)]--;
}
}
TrackedMetalBuffer(TrackedMetalBuffer&&) = delete;
TrackedMetalBuffer(TrackedMetalBuffer const&) = delete;
TrackedMetalBuffer& operator=(TrackedMetalBuffer const&) = delete;
TrackedMetalBuffer& operator=(TrackedMetalBuffer&& rhs) noexcept {
swap(rhs);
return *this;
}
id<MTLBuffer> get() const noexcept { return mBuffer; }
operator bool() const noexcept { return bool(mBuffer); }
static uint64_t getAliveBuffers() {
uint64_t sum = 0;
for (size_t i = 1; i < TypeCount; i++) {
sum += getAliveBuffers(static_cast<Type>(i));
for (const auto& v : aliveBuffers) {
sum += v;
}
return sum;
}
static uint64_t getAliveBuffers(Type type) {
assert_invariant(type != Type::NONE);
NSHashTable* hashTable = aliveBuffers[toIndex(type)];
// Caution! We can't simply use hashTable.count here, which is inaccurate.
// See http://cocoamine.net/blog/2013/12/13/nsmaptable-and-zeroing-weak-references/
return hashTable.objectEnumerator.allObjects.count;
return aliveBuffers[toIndex(type)];
}
#else
static void initialize() {}
static void setPlatform(MetalPlatform* p) {}
static id<MTLBuffer> track(id<MTLBuffer> buffer, Type type) { return buffer; }
static uint64_t getAliveBuffers() { return 0; }
static uint64_t getAliveBuffers(Type type) { return 0; }
#endif
static void setPlatform(MetalPlatform* p) { platform = p; }
private:
#if FILAMENT_METAL_BUFFER_TRACKING
static std::array<NSHashTable<id<MTLBuffer>>*, TypeCount> aliveBuffers;
void swap(TrackedMetalBuffer& other) noexcept {
std::swap(mBuffer, other.mBuffer);
std::swap(mType, other.mType);
}
id<MTLBuffer> mBuffer;
Type mType = Type::NONE;
static MetalPlatform* platform;
#endif
static std::array<uint64_t, TypeCount> aliveBuffers;
};
class MetalBuffer {
@@ -198,7 +204,7 @@ public:
private:
id<MTLBuffer> mBuffer;
TrackedMetalBuffer mBuffer;
size_t mBufferSize = 0;
void* mCpuBuffer = nullptr;
MetalContext& mContext;
@@ -247,11 +253,9 @@ public:
mBufferOptions(options),
mSlotSizeBytes(computeSlotSize(layout)),
mSlotCount(slotCount) {
{
ScopedAllocationTimer timer("ring");
mBuffer = [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions];
}
MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::RING);
ScopedAllocationTimer timer("ring");
mBuffer = { [device newBufferWithLength:mSlotSizeBytes * mSlotCount options:mBufferOptions],
TrackedMetalBuffer::Type::RING };
assert_invariant(mBuffer);
}
@@ -271,11 +275,11 @@ public:
// finishes executing.
{
ScopedAllocationTimer timer("ring");
mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
mAuxBuffer = { [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions],
TrackedMetalBuffer::Type::RING };
}
MetalBufferTracking::track(mAuxBuffer, MetalBufferTracking::Type::RING);
assert_invariant(mAuxBuffer);
return { mAuxBuffer, 0 };
return { mAuxBuffer.get(), 0 };
}
mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -304,9 +308,9 @@ public:
*/
std::pair<id<MTLBuffer>, NSUInteger> getCurrentAllocation() const {
if (UTILS_UNLIKELY(mAuxBuffer)) {
return { mAuxBuffer, 0 };
return { mAuxBuffer.get(), 0 };
}
return { mBuffer, mCurrentSlot * mSlotSizeBytes };
return { mBuffer.get(), mCurrentSlot * mSlotSizeBytes };
}
bool canAccomodateLayout(MTLSizeAndAlign layout) const {
@@ -315,8 +319,8 @@ public:
private:
id<MTLDevice> mDevice;
id<MTLBuffer> mBuffer;
id<MTLBuffer> mAuxBuffer;
TrackedMetalBuffer mBuffer;
TrackedMetalBuffer mAuxBuffer;
MTLResourceOptions mBufferOptions;

20
filament/backend/src/metal/MetalBuffer.mm
View File

@@ -22,14 +22,10 @@
namespace filament {
namespace backend {
std::array<uint64_t, TrackedMetalBuffer::TypeCount> TrackedMetalBuffer::aliveBuffers = { 0 };
MetalPlatform* TrackedMetalBuffer::platform = nullptr;
MetalPlatform* ScopedAllocationTimer::platform = nullptr;
#if FILAMENT_METAL_BUFFER_TRACKING
std::array<NSHashTable<id<MTLBuffer>>*, MetalBufferTracking::TypeCount>
MetalBufferTracking::aliveBuffers;
MetalPlatform* MetalBufferTracking::platform = nullptr;
#endif
MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
// If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
@@ -45,9 +41,9 @@ MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType,
// Otherwise, we allocate a private GPU buffer.
{
ScopedAllocationTimer timer("generic");
mBuffer = [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate];
mBuffer = { [context.device newBufferWithLength:size options:MTLResourceStorageModePrivate],
TrackedMetalBuffer::Type::GENERIC };
}
MetalBufferTracking::track(mBuffer, MetalBufferTracking::Type::GENERIC);
ASSERT_POSTCONDITION(mBuffer, "Could not allocate Metal buffer of size %zu.", size);
}
@@ -74,7 +70,7 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
// Acquire a staging buffer to hold the contents of this update.
MetalBufferPool* bufferPool = mContext.bufferPool;
const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
memcpy(staging->buffer.contents, src, size);
memcpy(staging->buffer.get().contents, src, size);
// The blit below requires that byteOffset be a multiple of 4.
ASSERT_PRECONDITION(!(byteOffset & 0x3u), "byteOffset must be a multiple of 4");
@@ -83,9 +79,9 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
blitEncoder.label = @"Buffer upload blit";
[blitEncoder copyFromBuffer:staging->buffer
[blitEncoder copyFromBuffer:staging->buffer.get()
sourceOffset:0
toBuffer:mBuffer
toBuffer:mBuffer.get()
destinationOffset:byteOffset
size:size];
[blitEncoder endEncoding];
@@ -106,7 +102,7 @@ id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) n
return nil;
}
assert_invariant(mBuffer);
return mBuffer;
return mBuffer.get();
}
void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncoder> encoder,

2
filament/backend/src/metal/MetalBufferPool.h
View File

@@ -32,7 +32,7 @@ struct MetalContext;
// Immutable POD representing a shared CPU-GPU buffer.
struct MetalBufferPoolEntry {
id<MTLBuffer> buffer;
TrackedMetalBuffer buffer;
size_t capacity;
mutable uint64_t lastAccessed;
mutable uint32_t referenceCount;

3
filament/backend/src/metal/MetalBufferPool.mm
View File

@@ -48,10 +48,9 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
buffer = [mContext.device newBufferWithLength:numBytes
options:MTLResourceStorageModeShared];
}
MetalBufferTracking::track(buffer, MetalBufferTracking::Type::STAGING);
ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
.buffer = buffer,
.buffer = { buffer, TrackedMetalBuffer::Type::STAGING },
.capacity = numBytes,
.lastAccessed = mCurrentFrame,
.referenceCount = 1

22
filament/backend/src/metal/MetalDriver.mm
View File

@@ -106,9 +106,8 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
mStereoscopicType(driverConfig.stereoscopicType) {
mContext->driver = this;
TrackedMetalBuffer::setPlatform(platform);
ScopedAllocationTimer::setPlatform(platform);
MetalBufferTracking::initialize();
MetalBufferTracking::setPlatform(platform);
mContext->device = mPlatform.createDevice();
assert_invariant(mContext->device);
@@ -203,7 +202,7 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
}
MetalDriver::~MetalDriver() noexcept {
MetalBufferTracking::setPlatform(nullptr);
TrackedMetalBuffer::setPlatform(nullptr);
ScopedAllocationTimer::setPlatform(nullptr);
mContext->device = nil;
mContext->emptyTexture = nil;
@@ -225,16 +224,13 @@ void MetalDriver::beginFrame(int64_t monotonic_clock_ns,
os_signpost_interval_begin(mContext->log, mContext->signpostId, "Frame encoding", "%{public}d", frameId);
#endif
if (mPlatform.hasDebugUpdateStatFunc()) {
#if FILAMENT_METAL_BUFFER_TRACKING
const uint64_t generic = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::GENERIC);
const uint64_t ring = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::RING);
const uint64_t staging = MetalBufferTracking::getAliveBuffers(MetalBufferTracking::Type::STAGING);
const uint64_t total = generic + ring + staging;
mPlatform.debugUpdateStat("filament.metal.alive_buffers", total);
mPlatform.debugUpdateStat("filament.metal.alive_buffers.generic", generic);
mPlatform.debugUpdateStat("filament.metal.alive_buffers.ring", ring);
mPlatform.debugUpdateStat("filament.metal.alive_buffers.staging", staging);
#endif
mPlatform.debugUpdateStat("filament.metal.alive_buffers", TrackedMetalBuffer::getAliveBuffers());
mPlatform.debugUpdateStat("filament.metal.alive_buffers.generic",
TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::GENERIC));
mPlatform.debugUpdateStat("filament.metal.alive_buffers.ring",
TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::RING));
mPlatform.debugUpdateStat("filament.metal.alive_buffers.staging",
TrackedMetalBuffer::getAliveBuffers(TrackedMetalBuffer::Type::STAGING));
}
}

4
filament/backend/src/metal/MetalHandles.mm
View File

@@ -800,13 +800,13 @@ void MetalTexture::loadWithCopyBuffer(uint32_t level, uint32_t slice, MTLRegion
PixelBufferDescriptor const& data, const PixelBufferShape& shape) {
const size_t stagingBufferSize = shape.totalBytes;
auto entry = context.bufferPool->acquireBuffer(stagingBufferSize);
memcpy(entry->buffer.contents,
memcpy(entry->buffer.get().contents,
static_cast<uint8_t*>(data.buffer) + shape.sourceOffset,
stagingBufferSize);
id<MTLCommandBuffer> blitCommandBuffer = getPendingCommandBuffer(&context);
id<MTLBlitCommandEncoder> blitCommandEncoder = [blitCommandBuffer blitCommandEncoder];
blitCommandEncoder.label = @"Texture upload buffer blit";
[blitCommandEncoder copyFromBuffer:entry->buffer
[blitCommandEncoder copyFromBuffer:entry->buffer.get()
sourceOffset:0
sourceBytesPerRow:shape.bytesPerRow
sourceBytesPerImage:shape.bytesPerSlice