P4 to Git Change 1576247 by gandryey@gera-w8 on 2018/07/04 16:55:54
SWDEV-158017 - CL_DEVICE_GLOBAL_FREE_MEMORY_AMD doesn't work correctly on PAL backend - Change the free memory detection to account possible memory allocation with oversubscription. - Make sure resource cache isn't counted as allocated memory Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#593 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#244 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.hpp#88 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#94 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#31 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#68 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#25 edit
This commit is contained in:
@@ -1805,7 +1805,8 @@ bool Device::globalFreeMemory(size_t* freeMemory) const {
|
||||
|
||||
// Fill free memory info
|
||||
freeMemory[TotalFreeMemory] =
|
||||
(memInfo.cardMemAvailableBytes + memInfo.cardExtMemAvailableBytes) / Ki;
|
||||
(memInfo.cardMemAvailableBytes + memInfo.cardExtMemAvailableBytes +
|
||||
resourceCache().lclCacheSize()) / Ki;
|
||||
freeMemory[LargestFreeBlock] =
|
||||
std::max(memInfo.cardLargestFreeBlockBytes, memInfo.cardExtLargestFreeBlockBytes) / Ki;
|
||||
if (settings().apuSystem_) {
|
||||
|
||||
@@ -1748,6 +1748,9 @@ bool ResourceCache::addCalResource(Resource::CalResourceDesc* desc, GslResourceR
|
||||
// Add the current resource to the cache
|
||||
resCache_.push_front({descCached, ref});
|
||||
cacheSize_ += size;
|
||||
if (desc->type_ == Resource::Local) {
|
||||
lclCacheSize_ += size;
|
||||
}
|
||||
result = true;
|
||||
}
|
||||
}
|
||||
@@ -1777,10 +1780,13 @@ GslResourceReference* ResourceCache::findCalResource(Resource::CalResourceDesc*
|
||||
(entry->mipLevels_ == desc->mipLevels_) && (entry->isAllocSVM_ == desc->isAllocSVM_) &&
|
||||
(entry->isAllocExecute_ == desc->isAllocExecute_)) {
|
||||
ref = it.second;
|
||||
cacheSize_ -= size;
|
||||
if (entry->type_ == Resource::Local) {
|
||||
lclCacheSize_ -= size;
|
||||
}
|
||||
delete it.first;
|
||||
// Remove the found etry from the cache
|
||||
resCache_.remove(it);
|
||||
cacheSize_ -= size;
|
||||
break;
|
||||
}
|
||||
}
|
||||
@@ -1823,12 +1829,16 @@ void ResourceCache::removeLast() {
|
||||
|
||||
size_t size = getResourceSize(entry.first);
|
||||
|
||||
cacheSize_ -= size;
|
||||
if (entry.first->type_ == Resource::Local) {
|
||||
lclCacheSize_ -= size;
|
||||
}
|
||||
|
||||
// Delete CalResourceDesc
|
||||
delete entry.first;
|
||||
|
||||
// Destroy GSL resource
|
||||
entry.second->release();
|
||||
cacheSize_ -= size;
|
||||
}
|
||||
|
||||
} // namespace gpu
|
||||
|
||||
@@ -428,7 +428,10 @@ class ResourceCache : public amd::HeapObject {
|
||||
public:
|
||||
//! Default constructor
|
||||
ResourceCache(size_t cacheSizeLimit)
|
||||
: lockCacheOps_("CAL resource cache", true), cacheSize_(0), cacheSizeLimit_(cacheSizeLimit) {}
|
||||
: lockCacheOps_("CAL resource cache", true)
|
||||
, cacheSize_(0)
|
||||
, lclCacheSize_(0)
|
||||
, cacheSizeLimit_(cacheSizeLimit) {}
|
||||
|
||||
//! Default destructor
|
||||
~ResourceCache();
|
||||
@@ -446,6 +449,9 @@ class ResourceCache : public amd::HeapObject {
|
||||
//! Destroys cache
|
||||
bool free(size_t minCacheEntries = 0);
|
||||
|
||||
//! Returns the size of local memory, stored in the cache
|
||||
size_t lclCacheSize() const { return lclCacheSize_; }
|
||||
|
||||
private:
|
||||
//! Disable copy constructor
|
||||
ResourceCache(const ResourceCache&);
|
||||
@@ -462,6 +468,7 @@ class ResourceCache : public amd::HeapObject {
|
||||
amd::Monitor lockCacheOps_; //!< Lock to serialise cache access
|
||||
|
||||
size_t cacheSize_; //!< Current cache size in bytes
|
||||
size_t lclCacheSize_; //!< Local memory stored in the cache
|
||||
size_t cacheSizeLimit_; //!< Cache size limit in bytes
|
||||
|
||||
//! CAL resource cache
|
||||
|
||||
@@ -846,15 +846,14 @@ bool Device::create(Pal::IDevice* device) {
|
||||
// Commit the new settings for the device
|
||||
result = iDev()->CommitSettingsAndInit();
|
||||
|
||||
Pal::GpuMemoryHeapProperties heaps[Pal::GpuHeapCount];
|
||||
iDev()->GetGpuMemoryHeapProperties(heaps);
|
||||
iDev()->GetGpuMemoryHeapProperties(heaps_);
|
||||
|
||||
Pal::WorkStationCaps wscaps = {};
|
||||
iDev()->QueryWorkStationCaps(&wscaps);
|
||||
|
||||
pal::Settings* gpuSettings = reinterpret_cast<pal::Settings*>(settings_);
|
||||
if ((gpuSettings == nullptr) ||
|
||||
!gpuSettings->create(properties(), heaps, wscaps, appProfile_.reportAsOCL12Device())) {
|
||||
!gpuSettings->create(properties(), heaps_, wscaps, appProfile_.reportAsOCL12Device())) {
|
||||
return false;
|
||||
}
|
||||
numComputeEngines_ = std::min(numComputeEngines_, settings().numComputeRings_);
|
||||
@@ -885,7 +884,7 @@ bool Device::create(Pal::IDevice* device) {
|
||||
}
|
||||
|
||||
// Fill the device info structure
|
||||
fillDeviceInfo(properties(), heaps, 16 * Ki, numComputeEngines(), numExclusiveComputeEngines());
|
||||
fillDeviceInfo(properties(), heaps_, 16 * Ki, numComputeEngines(), numExclusiveComputeEngines());
|
||||
|
||||
#ifdef DEBUG
|
||||
std::stringstream message;
|
||||
@@ -901,7 +900,7 @@ bool Device::create(Pal::IDevice* device) {
|
||||
#endif // DEBUG
|
||||
|
||||
for (uint i = 0; i < Pal::GpuHeap::GpuHeapCount; ++i) {
|
||||
freeMem[i] = heaps[i].heapSize;
|
||||
allocedMem[i] = 0;
|
||||
}
|
||||
|
||||
// Allocate SRD manager
|
||||
@@ -1677,15 +1676,31 @@ bool Device::globalFreeMemory(size_t* freeMemory) const {
|
||||
return false;
|
||||
}
|
||||
|
||||
Pal::gpusize local = freeMem[Pal::GpuHeapLocal];
|
||||
Pal::gpusize invisible = freeMem[Pal::GpuHeapInvisible];
|
||||
Pal::gpusize local = allocedMem[Pal::GpuHeapLocal];
|
||||
Pal::gpusize invisible = allocedMem[Pal::GpuHeapInvisible] - resourceCache().lclCacheSize();
|
||||
// Calculate free memory
|
||||
if (local >= heaps_[Pal::GpuHeapLocal].heapSize) {
|
||||
local = 0;
|
||||
} else {
|
||||
local = heaps_[Pal::GpuHeapLocal].heapSize - local;
|
||||
}
|
||||
if (invisible >= info().maxMemAllocSize_) {
|
||||
invisible = 0;
|
||||
} else {
|
||||
invisible = info().maxMemAllocSize_ - invisible;
|
||||
}
|
||||
|
||||
// Fill free memory info
|
||||
freeMemory[TotalFreeMemory] = static_cast<size_t>((local + invisible) / Ki);
|
||||
freeMemory[LargestFreeBlock] = static_cast<size_t>(std::max(local, invisible) / Ki);
|
||||
|
||||
if (settings().apuSystem_) {
|
||||
Pal::gpusize uswc = freeMem[Pal::GpuHeapGartUswc];
|
||||
Pal::gpusize uswc = allocedMem[Pal::GpuHeapGartUswc];
|
||||
if (uswc >= heaps_[Pal::GpuHeapGartUswc].heapSize) {
|
||||
uswc = 0;
|
||||
} else {
|
||||
uswc = heaps_[Pal::GpuHeapGartUswc].heapSize - uswc;
|
||||
}
|
||||
uswc /= Ki;
|
||||
freeMemory[TotalFreeMemory] += static_cast<size_t>(uswc);
|
||||
if (freeMemory[LargestFreeBlock] < uswc) {
|
||||
@@ -2124,11 +2139,11 @@ void Device::SrdManager::freeSrdSlot(uint64_t addr) {
|
||||
assert(false && "Wrong slot address!");
|
||||
}
|
||||
|
||||
void Device::updateFreeMemory(Pal::GpuHeap heap, Pal::gpusize size, bool free) {
|
||||
void Device::updateAllocedMemory(Pal::GpuHeap heap, Pal::gpusize size, bool free) const {
|
||||
if (free) {
|
||||
freeMem[heap] += size;
|
||||
} else {
|
||||
freeMem[heap] -= size;
|
||||
allocedMem[heap] -= size;
|
||||
} else {
|
||||
allocedMem[heap] += size;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -468,10 +468,10 @@ class Device : public NullDevice {
|
||||
RgpCaptureMgr* rgpCaptureMgr() const { return rgpCaptureMgr_; }
|
||||
|
||||
//! Update free memory for OCL extension
|
||||
void updateFreeMemory(Pal::GpuHeap heap, //!< PAL GPU heap for update
|
||||
Pal::gpusize size, //!< Size of alocated/destroyed memory
|
||||
bool free //!< TRUE if runtime frees memory
|
||||
);
|
||||
void updateAllocedMemory(Pal::GpuHeap heap, //!< PAL GPU heap for update
|
||||
Pal::gpusize size, //!< Size of alocated/destroyed memory
|
||||
bool free //!< TRUE if runtime frees memory
|
||||
) const;
|
||||
|
||||
//! Create internal blit program
|
||||
bool createBlitProgram();
|
||||
@@ -594,9 +594,10 @@ class Device : public NullDevice {
|
||||
mutable bool freeCPUMem_; //!< flag to mark GPU free SVM CPU mem
|
||||
Pal::DeviceProperties properties_; //!< PAL device properties
|
||||
Pal::IDevice* device_; //!< PAL device object
|
||||
std::atomic<Pal::gpusize> freeMem[Pal::GpuHeap::GpuHeapCount]; //!< Free memory counter
|
||||
mutable std::atomic<Pal::gpusize> allocedMem[Pal::GpuHeap::GpuHeapCount]; //!< Free memory counter
|
||||
std::unordered_set<Resource*>* resourceList_; //!< Active resource list
|
||||
RgpCaptureMgr* rgpCaptureMgr_; //!< RGP capture manager
|
||||
Pal::GpuMemoryHeapProperties heaps_[Pal::GpuHeapCount]; //!< Information about heaps, returned from PAL
|
||||
};
|
||||
|
||||
/*@}*/} // namespace pal
|
||||
|
||||
@@ -51,8 +51,10 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
|
||||
return nullptr;
|
||||
}
|
||||
}
|
||||
// Update free memory size counters
|
||||
const_cast<Device&>(dev).updateFreeMemory(createInfo.heaps[0], createInfo.size, false);
|
||||
if (!createInfo.flags.sdiExternal) {
|
||||
// Update free memory size counters
|
||||
dev.updateAllocedMemory(createInfo.heaps[0], createInfo.size, false);
|
||||
}
|
||||
return memRef;
|
||||
}
|
||||
|
||||
@@ -75,7 +77,7 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
|
||||
}
|
||||
}
|
||||
// Update free memory size counters
|
||||
const_cast<Device&>(dev).updateFreeMemory(Pal::GpuHeap::GpuHeapGartCacheable, createInfo.size, false);
|
||||
dev.updateAllocedMemory(Pal::GpuHeap::GpuHeapGartCacheable, createInfo.size, false);
|
||||
return memRef;
|
||||
}
|
||||
|
||||
@@ -97,8 +99,7 @@ GpuMemoryReference* GpuMemoryReference::Create(const Device& dev,
|
||||
}
|
||||
}
|
||||
// Update free memory size counters
|
||||
const_cast<Device&>(dev).updateFreeMemory(Pal::GpuHeap::GpuHeapGartCacheable, createInfo.size,
|
||||
false);
|
||||
dev.updateAllocedMemory(Pal::GpuHeap::GpuHeapGartCacheable, createInfo.size, false);
|
||||
return memRef;
|
||||
}
|
||||
|
||||
@@ -180,6 +181,12 @@ GpuMemoryReference::~GpuMemoryReference() {
|
||||
iMem()->Unmap();
|
||||
}
|
||||
if (0 != iMem()) {
|
||||
if (!(iMem()->Desc().flags.isShared ||
|
||||
iMem()->Desc().flags.isExternal ||
|
||||
iMem()->Desc().flags.isExternPhys)) {
|
||||
// Update free memory size counters
|
||||
device_.updateAllocedMemory(iMem()->Desc().preferredHeap, iMem()->Desc().size, true);
|
||||
}
|
||||
iMem()->Destroy();
|
||||
gpuMem_ = nullptr;
|
||||
}
|
||||
@@ -288,34 +295,6 @@ Resource::Resource(const Device& gpuDev, size_t width, size_t height, size_t dep
|
||||
|
||||
// ================================================================================================
|
||||
Resource::~Resource() {
|
||||
Pal::GpuHeap heap = Pal::GpuHeapCount;
|
||||
switch (memoryType()) {
|
||||
case Persistent:
|
||||
heap = Pal::GpuHeapLocal;
|
||||
break;
|
||||
case RemoteUSWC:
|
||||
heap = Pal::GpuHeapGartUswc;
|
||||
break;
|
||||
case Pinned:
|
||||
case Remote:
|
||||
heap = Pal::GpuHeapGartCacheable;
|
||||
break;
|
||||
case Shader:
|
||||
case BusAddressable:
|
||||
case ExternalPhysical:
|
||||
// Fall through to process the memory allocation ...
|
||||
case Local:
|
||||
heap = Pal::GpuHeapInvisible;
|
||||
break;
|
||||
default:
|
||||
heap = Pal::GpuHeapLocal;
|
||||
break;
|
||||
}
|
||||
if ((memRef_ != nullptr) && (heap != Pal::GpuHeapCount)) {
|
||||
// Update free memory size counters
|
||||
const_cast<Device&>(dev()).updateFreeMemory(heap, iMem()->Desc().size, true);
|
||||
}
|
||||
|
||||
free();
|
||||
|
||||
if ((nullptr != image_) &&
|
||||
@@ -1228,6 +1207,7 @@ void Resource::free()
|
||||
|
||||
// Add resource to the cache
|
||||
if (!dev().resourceCache().addGpuMemory(&desc_, memRef_, subOffset_)) {
|
||||
// Free PAL resource
|
||||
palFree();
|
||||
}
|
||||
}
|
||||
@@ -2015,6 +1995,9 @@ bool ResourceCache::addGpuMemory(Resource::Descriptor* desc,
|
||||
resCache_.push_front({descCached, ref});
|
||||
ref->gpu_ = nullptr;
|
||||
cacheSize_ += size;
|
||||
if (desc->type_ == Resource::Local) {
|
||||
lclCacheSize_ += size;
|
||||
}
|
||||
result = true;
|
||||
}
|
||||
}
|
||||
@@ -2056,10 +2039,13 @@ GpuMemoryReference* ResourceCache::findGpuMemory(Resource::Descriptor* desc, Pal
|
||||
(size > (sizeRes >> 2)) && ((it.second->iMem()->Desc().gpuVirtAddr % alignment) == 0) &&
|
||||
(entry->isAllocExecute_ == desc->isAllocExecute_)) {
|
||||
ref = it.second;
|
||||
cacheSize_ -= sizeRes;
|
||||
if (entry->type_ == Resource::Local) {
|
||||
lclCacheSize_ -= sizeRes;
|
||||
}
|
||||
delete it.first;
|
||||
// Remove the found etry from the cache
|
||||
resCache_.remove(it);
|
||||
cacheSize_ -= sizeRes;
|
||||
break;
|
||||
}
|
||||
}
|
||||
@@ -2090,9 +2076,12 @@ void ResourceCache::removeLast()
|
||||
amd::ScopedLock l(&lockCacheOps_);
|
||||
entry = resCache_.back();
|
||||
resCache_.pop_back();
|
||||
cacheSize_ -= entry.second->iMem()->Desc().size;
|
||||
if (entry.first->type_ == Resource::Local) {
|
||||
lclCacheSize_ -= entry.second->iMem()->Desc().size;
|
||||
}
|
||||
// Delete Descriptor
|
||||
delete entry.first;
|
||||
cacheSize_ -= entry.second->iMem()->Desc().size;
|
||||
}
|
||||
|
||||
// Destroy PAL resource
|
||||
|
||||
@@ -531,6 +531,7 @@ class ResourceCache : public amd::HeapObject {
|
||||
ResourceCache(Device* device, size_t cacheSizeLimit)
|
||||
: lockCacheOps_("PAL resource cache", true)
|
||||
, cacheSize_(0)
|
||||
, lclCacheSize_(0)
|
||||
, cacheSizeLimit_(cacheSizeLimit)
|
||||
, mem_sub_alloc_local_(device)
|
||||
, mem_sub_alloc_coarse_ (device)
|
||||
@@ -557,6 +558,9 @@ class ResourceCache : public amd::HeapObject {
|
||||
//! Returns true if cache was freed and false if cache is already empty.
|
||||
bool free(size_t minCacheEntries = 0);
|
||||
|
||||
//! Returns the size of local memory, stored in the cache
|
||||
size_t lclCacheSize() const { return lclCacheSize_; }
|
||||
|
||||
private:
|
||||
//! Disable copy constructor
|
||||
ResourceCache(const ResourceCache&);
|
||||
@@ -570,6 +574,7 @@ class ResourceCache : public amd::HeapObject {
|
||||
amd::Monitor lockCacheOps_; //!< Lock to serialise cache access
|
||||
|
||||
size_t cacheSize_; //!< Current cache size in bytes
|
||||
size_t lclCacheSize_; //!< Local memory stored in the cache
|
||||
const size_t cacheSizeLimit_; //!< Cache size limit in bytes
|
||||
|
||||
//! PAL resource cache
|
||||
|
||||
Reference in New Issue
Block a user