P4 to Git Change 1465654 by wchau@wchau_OCL_boltzmann on 2017/10/02 16:13:43
SWDEV-120036 - Supporting the cl_amd_device_attribute_query on the ROC device
- resubmit the codes with the addition of global free memory query support
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#66 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#24 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocmemory.cpp#26 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.cpp#27 edit
[ROCm/clr commit: 676a9d5003]
Этот коммит содержится в:
@@ -129,6 +129,7 @@ Device::Device(hsa_agent_t bkendDevice)
|
||||
, xferWrite_(nullptr)
|
||||
, pro_device_(nullptr)
|
||||
, pro_ena_(false)
|
||||
, freeMem_(0)
|
||||
, numOfVgpus_(0) {
|
||||
group_segment_.handle = 0;
|
||||
system_segment_.handle = 0;
|
||||
@@ -887,6 +888,8 @@ bool Device::populateOCLDeviceConstants() {
|
||||
}
|
||||
}
|
||||
|
||||
freeMem_ = info_.globalMemSize_;
|
||||
|
||||
// Make sure the max allocation size is not larger than the available
|
||||
// memory size.
|
||||
info_.maxMemAllocSize_ = std::min(info_.maxMemAllocSize_, info_.globalMemSize_);
|
||||
@@ -1093,26 +1096,35 @@ bool Device::populateOCLDeviceConstants() {
|
||||
#endif // !defined(WITH_LIGHTNING_COMPILER)
|
||||
}
|
||||
|
||||
//if (settings().checkExtension(ClAmdDeviceAttributeQuery)) {
|
||||
//info_.simdPerCU_ = deviceInfo_.simdPerCU_;
|
||||
//info_.simdWidth_ = deviceInfo_.simdWidth_;
|
||||
//info_.simdInstructionWidth_ = deviceInfo_.simdInstructionWidth_;
|
||||
if (settings().checkExtension(ClAmdDeviceAttributeQuery)) {
|
||||
info_.simdPerCU_ = deviceInfo_.simdPerCU_;
|
||||
info_.simdWidth_ = deviceInfo_.simdWidth_;
|
||||
info_.simdInstructionWidth_ = deviceInfo_.simdInstructionWidth_;
|
||||
if (HSA_STATUS_SUCCESS !=
|
||||
hsa_agent_get_info(_bkendDevice, HSA_AGENT_INFO_WAVEFRONT_SIZE, &info_.wavefrontWidth_)) {
|
||||
return false;
|
||||
}
|
||||
//info_.globalMemChannels_ = palProp.gpuMemoryProperties.performance.vramBusBitWidth / 32;
|
||||
//info_.globalMemChannelBanks_ = 4;
|
||||
//info_.globalMemChannelBankWidth_ = deviceInfo_.memChannelBankWidth_;
|
||||
//info_.localMemSizePerCU_ = deviceInfo_.localMemSizePerCU_;
|
||||
//info_.localMemBanks_ = deviceInfo_.localMemBanks_;
|
||||
if (HSA_STATUS_SUCCESS !=
|
||||
hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_MEMORY_WIDTH, &info_.globalMemChannels_)) {
|
||||
return false;
|
||||
}
|
||||
info_.globalMemChannelBanks_ = 4;
|
||||
info_.globalMemChannelBankWidth_ = deviceInfo_.memChannelBankWidth_;
|
||||
info_.localMemSizePerCU_ = deviceInfo_.localMemSizePerCU_;
|
||||
info_.localMemBanks_ = deviceInfo_.localMemBanks_;
|
||||
info_.gfxipVersion_ = deviceInfo_.gfxipVersion_;
|
||||
//info_.numAsyncQueues_ = numComputeRings;
|
||||
//info_.numRTQueues_ = numExclusiveComputeRings;
|
||||
//info_.numRTCUs_ = palProp.engineProperties[Pal::EngineTypeExclusiveCompute].maxNumDedicatedCu;
|
||||
//info_.threadTraceEnable_ = settings().threadTraceEnable_;
|
||||
//}
|
||||
|
||||
if (HSA_STATUS_SUCCESS !=
|
||||
hsa_agent_get_info(_bkendDevice, HSA_AGENT_INFO_QUEUES_MAX, &info_.numAsyncQueues_)) {
|
||||
return false;
|
||||
}
|
||||
info_.numRTQueues_ = info_.numAsyncQueues_;
|
||||
if (HSA_STATUS_SUCCESS !=
|
||||
hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT, &info_.numRTCUs_)) {
|
||||
return false;
|
||||
}
|
||||
//TODO: set to true once thread trace support is available
|
||||
info_.threadTraceEnable_ = false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
@@ -1136,7 +1148,18 @@ device::VirtualDevice* Device::createVirtualDevice(amd::CommandQueue* queue) {
|
||||
return virtualDevice;
|
||||
}
|
||||
|
||||
bool Device::globalFreeMemory(size_t* freeMemory) const { return false; }
|
||||
bool Device::globalFreeMemory(size_t* freeMemory) const {
|
||||
const uint TotalFreeMemory = 0;
|
||||
const uint LargestFreeBlock = 1;
|
||||
|
||||
freeMemory[TotalFreeMemory] = freeMem_ / Ki;
|
||||
|
||||
// since there is no memory heap on ROCm, the biggest free block is
|
||||
// equal to total free local memory
|
||||
freeMemory[LargestFreeBlock] = freeMemory[TotalFreeMemory];
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
bool Device::bindExternalDevice(uint flags, void* const gfxDevice[], void* gfxContext,
|
||||
bool validateOnly) {
|
||||
@@ -1399,6 +1422,15 @@ void Device::memFree(void* ptr, size_t size) const {
|
||||
}
|
||||
}
|
||||
|
||||
void Device::updateFreeMemory(size_t size, bool free) {
|
||||
if (free) {
|
||||
freeMem_ += size;
|
||||
}
|
||||
else {
|
||||
freeMem_ -= size;
|
||||
}
|
||||
}
|
||||
|
||||
void* Device::svmAlloc(amd::Context& context, size_t size, size_t alignment, cl_svm_mem_flags flags,
|
||||
void* svmPtr) const {
|
||||
amd::Memory* mem = nullptr;
|
||||
|
||||
@@ -400,6 +400,9 @@ class Device : public NullDevice {
|
||||
// Lock protect P2P staging operations
|
||||
const std::vector<Memory*>& P2PStages() const { return p2p_stages_; }
|
||||
|
||||
// Update the global free memory size
|
||||
void updateFreeMemory(size_t size, bool free);
|
||||
|
||||
private:
|
||||
static hsa_ven_amd_loader_1_00_pfn_t amd_loader_ext_table;
|
||||
|
||||
@@ -433,6 +436,7 @@ class Device : public NullDevice {
|
||||
XferBuffers* xferWrite_; //!< Transfer buffers write
|
||||
const IProDevice* pro_device_; //!< AMDGPUPro device
|
||||
bool pro_ena_; //!< Extra functionality with AMDGPUPro device, beyond ROCr
|
||||
std::atomic<size_t> freeMem_; //!< Total of free memory available
|
||||
|
||||
public:
|
||||
amd::Atomic<uint> numOfVgpus_; //!< Virtual gpu unique index
|
||||
|
||||
@@ -586,6 +586,7 @@ void Buffer::destroy() {
|
||||
}
|
||||
} else {
|
||||
dev().memFree(deviceMemory_, size());
|
||||
const_cast<Device&>(dev()).updateFreeMemory(size(), true);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -669,6 +670,9 @@ bool Buffer::create() {
|
||||
deviceMemory_ = dev().hostAlloc(size(), 1, false);
|
||||
owner()->setHostMem(deviceMemory_);
|
||||
}
|
||||
else {
|
||||
const_cast<Device&>(dev()).updateFreeMemory(size(), false);
|
||||
}
|
||||
|
||||
assert(amd::isMultipleOf(deviceMemory_, static_cast<size_t>(dev().info().memBaseAddrAlign_)));
|
||||
|
||||
@@ -932,6 +936,9 @@ bool Image::create() {
|
||||
if (originalDeviceMemory_ == nullptr) {
|
||||
originalDeviceMemory_ = dev().hostAlloc(alloc_size, 1, false);
|
||||
}
|
||||
else {
|
||||
const_cast<Device&>(dev()).updateFreeMemory(alloc_size, false);
|
||||
}
|
||||
|
||||
deviceMemory_ = reinterpret_cast<void*>(
|
||||
amd::alignUp(reinterpret_cast<uintptr_t>(originalDeviceMemory_), deviceImageInfo_.alignment));
|
||||
@@ -1092,6 +1099,7 @@ void Image::destroy() {
|
||||
|
||||
if (originalDeviceMemory_ != nullptr) {
|
||||
dev().memFree(originalDeviceMemory_, deviceImageInfo_.size);
|
||||
const_cast<Device&>(dev()).updateFreeMemory(size(), true);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -101,6 +101,8 @@ bool Settings::create(bool fullProfile, int gfxipVersion) {
|
||||
if (MesaInterop::Supported()) {
|
||||
enableExtension(ClKhrGlSharing);
|
||||
}
|
||||
// Enable platform extension
|
||||
enableExtension(ClAmdDeviceAttributeQuery);
|
||||
|
||||
// Enable KHR double precision extension
|
||||
enableExtension(ClKhrFp64);
|
||||
|
||||
Ссылка в новой задаче
Block a user