Ryujinx/moltenvk
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #1381 from js6i/volatile_buffer_locking

Browse Source 
MVKMTLBufferAllocator: Make sure temporary buffers stay mlocked() until they're encoded.
This commit is contained in:
Bill Hollings 2021-06-15 08:15:08 -04:00 committed by GitHub
commit d5b4ef1b19
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
10 changed files with 66 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
MoltenVK/MoltenVK/Commands/MVKCmdTransfer.mm
View File

@ -1612,11 +1612,9 @@ void MVKCmdUpdateBuffer::encode(MVKCommandEncoder* cmdEncoder) {
NSUInteger dstMTLBuffOffset = _dstBuffer->getMTLBufferOffset() + _dstOffset;
// Copy data to the source MTLBuffer
MVKMTLBufferAllocation* srcMTLBufferAlloc = (MVKMTLBufferAllocation*)cmdEncoder->getCommandEncodingPool()->acquireMTLBufferAllocation(_dataSize);
MVKMTLBufferAllocation* srcMTLBufferAlloc = cmdEncoder->getCommandEncodingPool()->acquireMTLBufferAllocation(_dataSize);
void* pBuffData = srcMTLBufferAlloc->getContents();
mlock(pBuffData, _dataSize);
memcpy(pBuffData, _srcDataCache.data(), _dataSize);
munlock(pBuffData, _dataSize);
[mtlBlitEnc copyFromBuffer: srcMTLBufferAlloc->_mtlBuffer
sourceOffset: srcMTLBufferAlloc->_offset

4
MoltenVK/MoltenVK/Commands/MVKCommandBuffer.mm
View File

@ -695,7 +695,7 @@ void MVKCommandEncoder::setComputeBytes(id<MTLComputeCommandEncoder> mtlEncoder,
// Return the MTLBuffer allocation to the pool once the command buffer is done with it
const MVKMTLBufferAllocation* MVKCommandEncoder::getTempMTLBuffer(NSUInteger length, bool isPrivate, bool isDedicated) {
const MVKMTLBufferAllocation* mtlBuffAlloc = getCommandEncodingPool()->acquireMTLBufferAllocation(length, isPrivate, isDedicated);
MVKMTLBufferAllocation* mtlBuffAlloc = getCommandEncodingPool()->acquireMTLBufferAllocation(length, isPrivate, isDedicated);
[_mtlCmdBuffer addCompletedHandler: ^(id<MTLCommandBuffer> mcb) { mtlBuffAlloc->returnToPool(); }];
return mtlBuffAlloc;
}
@ -708,9 +708,7 @@ MVKCommandEncodingPool* MVKCommandEncoder::getCommandEncodingPool() {
const MVKMTLBufferAllocation* MVKCommandEncoder::copyToTempMTLBufferAllocation(const void* bytes, NSUInteger length, bool isDedicated) {
const MVKMTLBufferAllocation* mtlBuffAlloc = getTempMTLBuffer(length, false, isDedicated);
void* pBuffData = mtlBuffAlloc->getContents();
mlock(pBuffData, length);
memcpy(pBuffData, bytes, length);
munlock(pBuffData, length);
return mtlBuffAlloc;
}

2
MoltenVK/MoltenVK/Commands/MVKCommandEncodingPool.h
View File

@ -66,7 +66,7 @@ public:
* To return the returned allocation back to the pool to be reused,
* call the returnToPool() function on the returned allocation.
*/
const MVKMTLBufferAllocation* acquireMTLBufferAllocation(NSUInteger length, bool isPrivate = false, bool isDedicated = false);
MVKMTLBufferAllocation* acquireMTLBufferAllocation(NSUInteger length, bool isPrivate = false, bool isDedicated = false);
/**
* Returns a MTLRenderPipelineState dedicated to rendering to several attachments

2
MoltenVK/MoltenVK/Commands/MVKCommandEncodingPool.mm
View File

@ -77,7 +77,7 @@ id<MTLDepthStencilState> MVKCommandEncodingPool::getMTLDepthStencilState(bool us
MVK_ENC_REZ_ACCESS(_cmdClearDefaultDepthStencilState, newMTLDepthStencilState(useDepth, useStencil));
}
const MVKMTLBufferAllocation* MVKCommandEncodingPool::acquireMTLBufferAllocation(NSUInteger length, bool isPrivate, bool isDedicated) {
MVKMTLBufferAllocation* MVKCommandEncodingPool::acquireMTLBufferAllocation(NSUInteger length, bool isPrivate, bool isDedicated) {
MVKAssert(isPrivate || !isDedicated, "Dedicated, host-shared temporary buffers are not supported.");
if (isDedicated) {
return _dedicatedMtlBufferAllocator.acquireMTLBufferRegion(length);

24
MoltenVK/MoltenVK/Commands/MVKMTLBufferAllocation.h
View File

@ -51,17 +51,20 @@ public:
MVKMTLBufferAllocationPool* getPool() const { return _pool; }
/** Returns this object back to the pool that created it. */
void returnToPool() const;
void returnToPool();
/** Constructs this instance with the specified pool as its origin. */
MVKMTLBufferAllocation(MVKMTLBufferAllocationPool* pool,
id<MTLBuffer> mtlBuffer,
NSUInteger offset,
NSUInteger length) : _pool(pool), _mtlBuffer(mtlBuffer), _offset(offset), _length(length) {}
NSUInteger length,
uint64_t poolIndex) : _pool(pool), _mtlBuffer(mtlBuffer), _offset(offset), _length(length), _poolIndex(poolIndex) {}
protected:
MVKMTLBufferAllocationPool* _pool;
friend class MVKMTLBufferAllocationPool;
MVKMTLBufferAllocationPool* _pool;
uint64_t _poolIndex;
};
@ -78,6 +81,11 @@ protected:
class MVKMTLBufferAllocationPool : public MVKObjectPool<MVKMTLBufferAllocation>, public MVKDeviceTrackingMixin {
public:
/** Returns a new allocation. */
MVKMTLBufferAllocation* acquireAllocation();
/** Returns a new allocation (without mutual exclusion). */
MVKMTLBufferAllocation* acquireAllocationUnlocked();
/** Returns the Vulkan API opaque object controlling this object. */
MVKVulkanAPIObject* getVulkanAPIObject() override { return _device->getVulkanAPIObject(); };
@ -93,7 +101,8 @@ protected:
MVKBaseObject* getBaseObject() override { return this; };
MVKMTLBufferAllocation* newObject() override;
void returnAllocation(MVKMTLBufferAllocation* ba) { _isThreadSafe ? returnObjectSafely(ba) : returnObject(ba); }
void returnAllocationUnlocked(MVKMTLBufferAllocation* ba);
void returnAllocation(MVKMTLBufferAllocation* ba);
uint32_t calcMTLBufferAllocationCount();
void addMTLBuffer();
@ -101,8 +110,9 @@ protected:
NSUInteger _allocationLength;
NSUInteger _mtlBufferLength;
MTLStorageMode _mtlStorageMode;
MVKSmallVector<id<MTLBuffer>, 64> _mtlBuffers;
bool _isThreadSafe;
struct MTLBufferTracker { id<MTLBuffer> mtlBuffer; uint64_t allocationCount; };
MVKSmallVector<MTLBufferTracker, 64> _mtlBuffers;
bool _isThreadSafe;
};
@ -132,7 +142,7 @@ public:
* To return the MVKMTLBufferAllocation back to the pool, call
* the returnToPool() function on the returned instance.
*/
const MVKMTLBufferAllocation* acquireMTLBufferRegion(NSUInteger length);
MVKMTLBufferAllocation* acquireMTLBufferRegion(NSUInteger length);
/**
* Configures this instance to dispense MVKMTLBufferAllocation up to the specified

53
MoltenVK/MoltenVK/Commands/MVKMTLBufferAllocation.mm
View File

@ -24,7 +24,7 @@
MVKVulkanAPIObject* MVKMTLBufferAllocation::getVulkanAPIObject() { return _pool->getVulkanAPIObject(); };
void MVKMTLBufferAllocation::returnToPool() const { _pool->returnAllocation((MVKMTLBufferAllocation*)this); }
void MVKMTLBufferAllocation::returnToPool() { _pool->returnAllocation(this); }
#pragma mark -
@ -39,16 +39,49 @@ MVKMTLBufferAllocation* MVKMTLBufferAllocationPool::newObject() {
// of future allocation to beyond this allocation.
NSUInteger offset = _nextOffset;
_nextOffset += _allocationLength;
return new MVKMTLBufferAllocation(this, _mtlBuffers.back(), offset, _allocationLength);
return new MVKMTLBufferAllocation(this, _mtlBuffers.back().mtlBuffer, offset, _allocationLength, _mtlBuffers.size() - 1);
}
// Adds a new MTLBuffer to the buffer pool and resets the next offset to the start of it
void MVKMTLBufferAllocationPool::addMTLBuffer() {
MTLResourceOptions mbOpts = (_mtlStorageMode << MTLResourceStorageModeShift) | MTLResourceCPUCacheModeDefaultCache;
_mtlBuffers.push_back([_device->getMTLDevice() newBufferWithLength: _mtlBufferLength options: mbOpts]);
_mtlBuffers.push_back({ [_device->getMTLDevice() newBufferWithLength: _mtlBufferLength options: mbOpts], 0 });
_nextOffset = 0;
}
MVKMTLBufferAllocation* MVKMTLBufferAllocationPool::acquireAllocationUnlocked() {
MVKMTLBufferAllocation* ba = acquireObject();
if (!_mtlBuffers[ba->_poolIndex].allocationCount++) {
[ba->_mtlBuffer setPurgeableState: MTLPurgeableStateNonVolatile];
}
return ba;
}
MVKMTLBufferAllocation* MVKMTLBufferAllocationPool::acquireAllocation() {
if (_isThreadSafe) {
std::lock_guard<std::mutex> lock(_lock);
return acquireAllocationUnlocked();
} else {
return acquireAllocationUnlocked();
}
}
void MVKMTLBufferAllocationPool::returnAllocationUnlocked(MVKMTLBufferAllocation* ba) {
if (!--_mtlBuffers[ba->_poolIndex].allocationCount) {
[ba->_mtlBuffer setPurgeableState: MTLPurgeableStateVolatile];
}
returnObject(ba);
}
void MVKMTLBufferAllocationPool::returnAllocation(MVKMTLBufferAllocation* ba) {
if (_isThreadSafe) {
std::lock_guard<std::mutex> lock(_lock);
returnAllocationUnlocked(ba);
} else {
returnAllocationUnlocked(ba);
}
}
MVKMTLBufferAllocationPool::MVKMTLBufferAllocationPool(MVKDevice* device, NSUInteger allocationLength, bool makeThreadSafe,
bool isDedicated, MTLStorageMode mtlStorageMode) :
@ -73,14 +106,17 @@ uint32_t MVKMTLBufferAllocationPool::calcMTLBufferAllocationCount() {
}
MVKMTLBufferAllocationPool::~MVKMTLBufferAllocationPool() {
mvkReleaseContainerContents(_mtlBuffers);
for (uint32_t bufferIndex = 0; bufferIndex < _mtlBuffers.size(); ++bufferIndex) {
[_mtlBuffers[bufferIndex].mtlBuffer release];
}
_mtlBuffers.clear();
}
#pragma mark -
#pragma mark MVKMTLBufferAllocator
const MVKMTLBufferAllocation* MVKMTLBufferAllocator::acquireMTLBufferRegion(NSUInteger length) {
MVKMTLBufferAllocation* MVKMTLBufferAllocator::acquireMTLBufferRegion(NSUInteger length) {
MVKAssert(length <= _maxAllocationLength, "This MVKMTLBufferAllocator has been configured to dispense MVKMTLBufferRegions no larger than %lu bytes.", (unsigned long)_maxAllocationLength);
// Can't allocate a segment smaller than the minimum MTLBuffer alignment.
@ -88,12 +124,7 @@ const MVKMTLBufferAllocation* MVKMTLBufferAllocator::acquireMTLBufferRegion(NSUI
// Convert max length to the next power-of-two exponent to use as a lookup
NSUInteger p2Exp = mvkPowerOfTwoExponent(length);
MVKMTLBufferAllocationPool* pRP = _regionPools[p2Exp];
const MVKMTLBufferAllocation* region = _isThreadSafe ? pRP->acquireObjectSafely() : pRP->acquireObject();
if (region) {
[region->_mtlBuffer setPurgeableState: MTLPurgeableStateVolatile];
}
return region;
return _regionPools[p2Exp]->acquireAllocation();
}
MVKMTLBufferAllocator::MVKMTLBufferAllocator(MVKDevice* device, NSUInteger maxRegionLength, bool makeThreadSafe, bool isDedicated, MTLStorageMode mtlStorageMode) : MVKBaseDeviceObject(device) {

2
MoltenVK/MoltenVK/GPUObjects/MVKDescriptor.h
View File

@ -397,7 +397,7 @@ public:
protected:
inline uint8_t* getData() { return _mvkMTLBufferAllocation ? (uint8_t*)_mvkMTLBufferAllocation->getContents() : nullptr; }
const MVKMTLBufferAllocation* _mvkMTLBufferAllocation = nullptr;
MVKMTLBufferAllocation* _mvkMTLBufferAllocation = nullptr;
};

2
MoltenVK/MoltenVK/GPUObjects/MVKDescriptorSet.h
View File

@ -180,7 +180,7 @@ public:
VkWriteDescriptorSetInlineUniformBlockEXT* pInlineUniformBlock);
/** Returns an MTLBuffer region allocation. */
const MVKMTLBufferAllocation* acquireMTLBufferRegion(NSUInteger length);
MVKMTLBufferAllocation* acquireMTLBufferRegion(NSUInteger length);
/**
* Returns the Metal argument buffer to which resources are written,
* or return nil if Metal argument buffers are not being used.

2
MoltenVK/MoltenVK/GPUObjects/MVKDescriptorSet.mm
View File

@ -339,7 +339,7 @@ void MVKDescriptorSet::read(const VkCopyDescriptorSet* pDescriptorCopy,
}
}
const MVKMTLBufferAllocation* MVKDescriptorSet::acquireMTLBufferRegion(NSUInteger length) {
MVKMTLBufferAllocation* MVKDescriptorSet::acquireMTLBufferRegion(NSUInteger length) {
return _pool->_inlineBlockMTLBufferAllocator.acquireMTLBufferRegion(length);
}

2
MoltenVK/MoltenVK/GPUObjects/MVKQueryPool.mm
View File

@ -223,9 +223,7 @@ id<MTLBuffer> MVKTimestampQueryPool::getResultBuffer(MVKCommandEncoder* cmdEncod
const MVKMTLBufferAllocation* tempBuff = cmdEncoder->getTempMTLBuffer(queryCount * _queryElementCount * sizeof(uint64_t));
void* pBuffData = tempBuff->getContents();
size_t size = queryCount * _queryElementCount * sizeof(uint64_t);
mlock(pBuffData, size);
memcpy(pBuffData, &_timestamps[firstQuery], size);
munlock(pBuffData, size);
offset = tempBuff->_offset;
return tempBuff->_mtlBuffer;
}