AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4 to Git Change 1781985 by gandryey@gera-win10 on 2019/05/13 11:54:40

Browse Source 
SWDEV-189140 - Add P2P support in PAL path
	- PAL requires P2P resource open on the usage device. Add the new interface to open the resource
	- Add a hidden P2P device object creation into amd::Memory. It can be activated with OCL context that has a single device.

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_p2p_amd.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#337 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#134 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#74 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#28 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#80 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.hpp#23 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#133 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#126 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#93 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.cpp#136 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/memory.hpp#109 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#306 edit
This commit is contained in:
foreman
2019-05-13 12:19:10 -04:00
parent 32135bf5cf
commit e16de03a2e
13 changed files with 306 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/runtime/platform/command.cpp
+8 -1
View File
@@ -552,7 +552,14 @@ bool CopyMemoryP2PCommand::validateMemory() {
LogPrintfError("Can't allocate memory size - 0x%08X bytes!", memory2_->getSize());
return false;
}
if (devices[0]->P2PStage() != nullptr) {
bool p2pStaging = false;
// Validate P2P memories on the current device, if any of them is null, then it's p2p staging
if ((nullptr == memory1_->getDeviceMemory(queue()->device())) ||
(nullptr == memory2_->getDeviceMemory(queue()->device()))) {
p2pStaging = true;
}
if (devices[0]->P2PStage() != nullptr && p2pStaging) {
amd::ScopedLock lock(devices[0]->P2PStageOps());
// Make sure runtime allocates memory on every device
for (uint d = 0; d < devices[0]->GlbCtx().devices().size(); ++d) {
rocclr/runtime/platform/memory.cpp
+40 -16
View File
@@ -70,14 +70,14 @@ Memory::Memory(Context& context, Type type, Flags flags, size_t size, void* svmP
version_(0),
lastWriter_(NULL),
interopObj_(NULL),
isParent_(false),
vDev_(NULL),
forceSysMemAlloc_(false),
mapCount_(0),
svmHostAddress_(svmPtr),
svmPtrCommited_(flags & CL_MEM_SVM_FINE_GRAIN_BUFFER ? true : false),
canBeCached_(true),
lockMemoryOps_("Memory Ops Lock", true) {}
flagsEx_(0),
lockMemoryOps_("Memory Ops Lock", true) {
svmPtrCommited_ = (flags & CL_MEM_SVM_FINE_GRAIN_BUFFER) ? true : false;
canBeCached_ = true;
}
Memory::Memory(Memory& parent, Flags flags, size_t origin, size_t size, Type type)
: numDevices_(0),
@@ -93,14 +93,13 @@ Memory::Memory(Memory& parent, Flags flags, size_t origin, size_t size, Type typ
version_(parent.getVersion()),
lastWriter_(parent.getLastWriter()),
interopObj_(parent.getInteropObj()),
isParent_(false),
vDev_(NULL),
forceSysMemAlloc_(false),
mapCount_(0),
svmHostAddress_(parent.getSvmPtr()),
svmPtrCommited_(parent.isSvmPtrCommited()),
canBeCached_(true),
flagsEx_(0),
lockMemoryOps_("Memory Ops Lock", true) {
svmPtrCommited_ = parent.isSvmPtrCommited();
canBeCached_ = true;
parent_->retain();
parent_->isParent_ = true;
@@ -126,13 +125,31 @@ Memory::Memory(Memory& parent, Flags flags, size_t origin, size_t size, Type typ
}
}
uint32_t Memory::NumDevicesWithP2P() {
uint32_t devices = context_().devices().size();
if (devices == 1) {
// Add p2p devices for allocation
devices = context_().devices().size() + context_().devices()[0]->P2PAccessDevices().size();
if (devices > 1) {
p2pAccess_ = true;
}
}
return devices;
}
void Memory::initDeviceMemory() {
deviceMemories_ = reinterpret_cast<DeviceMemory*>(reinterpret_cast<char*>(this) + sizeof(Memory));
memset(deviceMemories_, 0, context_().devices().size() * sizeof(DeviceMemory));
memset(deviceMemories_, 0, NumDevicesWithP2P() * sizeof(DeviceMemory));
}
void* Memory::operator new(size_t size, const Context& context) {
return RuntimeObject::operator new(size + context.devices().size() * sizeof(DeviceMemory));
uint32_t devices = context.devices().size();
if (devices == 1) {
// Add p2p devices for allocation
devices = context.devices().size() + context.devices()[0]->P2PAccessDevices().size();
}
return RuntimeObject::operator new(size + devices * sizeof(DeviceMemory));
}
void Memory::operator delete(void* p) { RuntimeObject::operator delete(p); }
@@ -259,7 +276,14 @@ bool Memory::addDeviceMemory(const Device* dev) {
bool result = false;
AllocState create = AllocCreate;
AllocState init = AllocInit;
if (make_atomic(deviceAlloced_[dev]).compareAndSet(init, create)) {
// Check if runtime already allocated all available slots for device memory
if (numDevices() == NumDevicesWithP2P()) {
// Mark the allocation as an empty
deviceAlloced_[dev] = AllocInit;
return false;
}
device::Memory* dm = dev->createMemory(*this);
// Add the new memory allocation to the device map
@@ -267,7 +291,7 @@ bool Memory::addDeviceMemory(const Device* dev) {
deviceMemories_[numDevices_].ref_ = dev;
deviceMemories_[numDevices_].value_ = dm;
numDevices_++;
assert((numDevices() <= context_().devices().size()) && "Too many device objects");
assert((numDevices() <= NumDevicesWithP2P()) && "Too many device objects");
// Mark the allocation with the complete flag
deviceAlloced_[dev] = AllocComplete;
@@ -275,6 +299,7 @@ bool Memory::addDeviceMemory(const Device* dev) {
svmBase_ = dm;
}
} else {
LogError("Video memory allocation failed!");
// Mark the allocation as an empty
deviceAlloced_[dev] = AllocInit;
}
@@ -322,7 +347,6 @@ device::Memory* Memory::getDeviceMemory(const Device& dev, bool alloc) {
if ((NULL == dm) && alloc) {
if (!addDeviceMemory(&dev)) {
LogError("Video memory allocation failed!");
return NULL;
}
dm = deviceMemories_[numDevices() - 1].value_;
@@ -446,7 +470,7 @@ void Memory::uncommitSvmMemory() {
void Buffer::initDeviceMemory() {
deviceMemories_ = reinterpret_cast<DeviceMemory*>(reinterpret_cast<char*>(this) + sizeof(Buffer));
memset(deviceMemories_, 0, context_().devices().size() * sizeof(DeviceMemory));
memset(deviceMemories_, 0, NumDevicesWithP2P() * sizeof(DeviceMemory));
}
bool Buffer::create(void* initFrom, bool sysMemAlloc, bool skipAlloc) {
@@ -472,7 +496,7 @@ bool Buffer::validateRegion(const Coord3D& origin, const Coord3D& region) const
void Pipe::initDeviceMemory() {
deviceMemories_ = reinterpret_cast<DeviceMemory*>(reinterpret_cast<char*>(this) + sizeof(Pipe));
memset(deviceMemories_, 0, context_().devices().size() * sizeof(DeviceMemory));
memset(deviceMemories_, 0, NumDevicesWithP2P() * sizeof(DeviceMemory));
}
#define GETMIPDIM(dim, mip) (((dim >> mip) > 0) ? (dim >> mip) : 1)
@@ -631,7 +655,7 @@ void Image::initDimension() {
void Image::initDeviceMemory() {
deviceMemories_ = reinterpret_cast<DeviceMemory*>(reinterpret_cast<char*>(this) + sizeof(Image));
memset(deviceMemories_, 0, context_().devices().size() * sizeof(DeviceMemory));
memset(deviceMemories_, 0, NumDevicesWithP2P() * sizeof(DeviceMemory));
}
bool Image::create(void* initFrom) { return Memory::create(initFrom); }
rocclr/runtime/platform/memory.hpp
+46 -31
View File
@@ -24,7 +24,7 @@
namespace device {
class Memory;
class VirtualDevice;
}
} // namespace device
namespace amd {
@@ -42,7 +42,7 @@ struct BufferRect : public amd::EmbeddedObject {
const size_t* region, //!< Copy region
size_t bufferRowPitch, //!< Provided buffer's row pitch
size_t bufferSlicePitch //!< Provided buffer's slice pitch
);
);
//! Returns the plain offset for the (X, Y, Z) location
size_t offset(size_t x, //!< Coordinate in X dimension
@@ -120,6 +120,9 @@ class Memory : public amd::RuntimeObject {
typedef cl_mem_flags Flags;
typedef DeviceMap<const Device*, device::Memory*> DeviceMemory;
//! Returns the number of devices this memory object is associated, including P2P access
uint32_t NumDevicesWithP2P();
size_t numDevices_; //!< Number of devices
//! The device memory objects included in this memory
@@ -141,13 +144,19 @@ class Memory : public amd::RuntimeObject {
size_t version_; //!< Update count, used for coherency
const Device* lastWriter_; //!< Which device wrote most recently (NULL if host)
InteropObject* interopObj_; //!< Interop object
bool isParent_; //!< This object is a parent
device::VirtualDevice* vDev_; //!< Memory object belongs to a virtual device only
bool forceSysMemAlloc_; //!< Forces system memory allocation
std::atomic_uint mapCount_; //!< Keep track of number of mappings for a memory object
void* svmHostAddress_; //!< svm host address;
bool svmPtrCommited_; //!< svm host address committed flag;
bool canBeCached_; //!< flag to if the object can be cached;
union {
struct {
uint32_t isParent_ : 1; //!< This object is a parent
uint32_t forceSysMemAlloc_ : 1; //!< Forces system memory allocation
uint32_t svmPtrCommited_ : 1; //!< svm host address committed flag
uint32_t canBeCached_ : 1; //!< flag to if the object can be cached
uint32_t p2pAccess_ : 1; //!< Memory object allows P2P access
};
uint32_t flagsEx_;
};
private:
//! Disable default assignment operator
@@ -168,20 +177,20 @@ class Memory : public amd::RuntimeObject {
Flags flags, //!< Object's flags
size_t size, //!< Memory size
void* svmPtr = NULL //!< svm host memory address, NULL if no SVM mem object
);
);
Memory(Memory& parent, //!< Context object
Flags flags, //!< Object's flags
size_t offset, //!< Memory offset
size_t size, //!< Memory size
Type type = 0 //!< Memory type
);
);
//! Memory object destructor
virtual ~Memory();
//! Copies initialization data to the backing store
virtual void copyToBackingStore(void* initFrom //!< Pointer to the initialization memory
);
);
//! Initializes the device memory array
virtual void initDeviceMemory();
@@ -193,14 +202,14 @@ class Memory : public amd::RuntimeObject {
//! Placement new operator.
void* operator new(size_t size, //!< Original allocation size
const Context& context //!< Context this memory object is allocated in.
);
);
// Provide a "matching" placement delete operator.
void operator delete(void*, //!< Pointer to deallocate
const Context& context //!< Context this memory object is allocated in.
);
);
// and a regular delete operator to satisfy synthesized methods.
void operator delete(void* //!< Pointer to deallocate
);
);
//! Returns the memory lock object
amd::Monitor& lockMemoryOps() { return lockMemoryOps_; }
@@ -228,32 +237,32 @@ class Memory : public amd::RuntimeObject {
virtual Pipe* asPipe() { return NULL; }
//! Creates and initializes device (cache) memory for all devices
virtual bool create(void* initFrom = NULL, //!< Pointer to the initialization data
bool sysMemAlloc = false, //!< Allocate device memory in system memory
bool skipAlloc = false //!< Skip device memory allocation
);
virtual bool create(void* initFrom = NULL, //!< Pointer to the initialization data
bool sysMemAlloc = false, //!< Allocate device memory in system memory
bool skipAlloc = false //!< Skip device memory allocation
);
//! Allocates device (cache) memory for a specific device
bool addDeviceMemory(const Device* dev //!< Device object
);
);
//! Replaces device (cache) memory for a specific device
void replaceDeviceMemory(const Device* dev, //!< Device object
device::Memory* dm //!< New device memory object for replacement
);
);
//! Find the section for the given device. Return NULL if not found.
device::Memory* getDeviceMemory(const Device& dev, //!< Device object
bool alloc = true //!< Allocates memory
);
);
//! Allocate host memory (as required)
bool allocHostMemory(void* initFrom, //!< Host memory provided by the application
bool allocHostMem, //!< Force system memory allocation
bool forceCopy = false //!< Force system memory allocation
);
);
virtual void IpcCreate (size_t offset, size_t* mem_size, void* handle) const {
virtual void IpcCreate(size_t offset, size_t* mem_size, void* handle) const {
ShouldNotReachHere();
}
@@ -314,8 +323,14 @@ class Memory : public amd::RuntimeObject {
void uncommitSvmMemory();
void setCacheStatus(bool canBeCached) {
canBeCached_ = canBeCached;
} //!< set the memobject cached status;
bool canBeCached() const { return canBeCached_; } //!< get the memobject cached status;
} //!< set the memobject cached status
bool canBeCached() const { return canBeCached_; } //!< get the memobject cached status
//! Check if this objects allows P2P access
bool P2PAccess() const { return p2pAccess_; }
//! Returns the base device memory object for possible P2P access
device::Memory* BaseP2PMemory() const { return deviceMemories_[0].value_; }
device::Memory* svmBase() const { return svmBase_; } //!< Returns SVM base for MGPU case
};
@@ -339,10 +354,10 @@ class Buffer : public Memory {
Buffer(Memory& parent, Flags flags, size_t origin, size_t size)
: Memory(parent, flags, origin, size) {}
bool create(void* initFrom = NULL, //!< Pointer to the initialization data
bool sysMemAlloc = false, //!< Allocate device memory in system memory
bool skipAlloc = false //!< Skip device memory allocation
);
bool create(void* initFrom = NULL, //!< Pointer to the initialization data
bool sysMemAlloc = false, //!< Allocate device memory in system memory
bool skipAlloc = false //!< Skip device memory allocation
);
//! static_cast to Buffer with sanity check
virtual Buffer* asBuffer() { return this; }
@@ -463,7 +478,7 @@ class Image : public Memory {
//! Copies initialization data to the backing store
virtual void copyToBackingStore(void* initFrom //!< Pointer to the initialization memory
);
);
void initDimension();
@@ -482,7 +497,7 @@ class Image : public Memory {
size_t height, //!< Image height
size_t depth, //!< Image depth
size_t arraySize //!< Image array size
);
);
const Format& getImageFormat() const { return impl_.format_; }
@@ -512,7 +527,7 @@ class Image : public Memory {
device::VirtualDevice* vDev, //!< Virtual device object
uint baseMipLevel = 0, //!< Base mip level for a view
cl_mem_flags flags = 0 //!< Memory allocation flags
);
);
//! Returns the impl for this image.
Impl& getImpl() { return impl_; }
@@ -554,7 +569,7 @@ class Image : public Memory {
//! Creates and initializes device (cache) memory for all devices
bool create(void* initFrom = NULL //!< Pointer to the initialization data
);
);
};
//! SVM-related functionality.