1a2158755SJunchao Zhang #include "../../interface/sycldevice.hpp" 2a2158755SJunchao Zhang #include <csetjmp> // for MPI sycl device awareness 3a2158755SJunchao Zhang #include <csignal> // SIGSEGV 4a2158755SJunchao Zhang #include <vector> 5a2158755SJunchao Zhang #include <CL/sycl.hpp> 6a2158755SJunchao Zhang 7a2158755SJunchao Zhang #if PetscDefined(USE_LOG) 8a2158755SJunchao Zhang PETSC_INTERN PetscErrorCode PetscLogInitialize(void); 9a2158755SJunchao Zhang #else 10a2158755SJunchao Zhang #define PetscLogInitialize() 0 11a2158755SJunchao Zhang #endif 12a2158755SJunchao Zhang 13a2158755SJunchao Zhang namespace Petsc 14a2158755SJunchao Zhang { 15a2158755SJunchao Zhang 1617f48955SJacob Faibussowitsch namespace Device 1717f48955SJacob Faibussowitsch { 1817f48955SJacob Faibussowitsch 1917f48955SJacob Faibussowitsch namespace SYCL 2017f48955SJacob Faibussowitsch { 2117f48955SJacob Faibussowitsch 22a2158755SJunchao Zhang // definition for static 2317f48955SJacob Faibussowitsch std::array<Device::DeviceInternal*,PETSC_DEVICE_MAX_DEVICES> Device::devices_array_ = {}; 2417f48955SJacob Faibussowitsch Device::DeviceInternal** Device::devices_ = &Device::devices_array_[1]; 2517f48955SJacob Faibussowitsch int Device::defaultDevice_ = PETSC_SYCL_DEVICE_NONE; 2617f48955SJacob Faibussowitsch bool Device::initialized_ = false; 27a2158755SJunchao Zhang 28a2158755SJunchao Zhang static std::jmp_buf MPISyclAwareJumpBuffer; 29a2158755SJunchao Zhang static bool MPISyclAwareJumpBufferSet; 30a2158755SJunchao Zhang 31a2158755SJunchao Zhang // internal "impls" class for SyclDevice. Each instance represents a single sycl device 3217f48955SJacob Faibussowitsch class PETSC_NODISCARD Device::DeviceInternal 33a2158755SJunchao Zhang { 34a2158755SJunchao Zhang const int id_; // -1 for the host device; 0 and up for gpu devices 35a2158755SJunchao Zhang bool devInitialized_; 36a2158755SJunchao Zhang const sycl::device syclDevice_; 37a2158755SJunchao Zhang 38a2158755SJunchao Zhang public: 39a2158755SJunchao Zhang // default constructor 4017f48955SJacob Faibussowitsch DeviceInternal(int id) noexcept : id_(id),devInitialized_(false),syclDevice_(chooseSYCLDevice_(id)){} 41a2158755SJunchao Zhang int id() const {return id_;} 42a2158755SJunchao Zhang bool initialized() const {return devInitialized_;} 43a2158755SJunchao Zhang 44a2158755SJunchao Zhang PETSC_NODISCARD PetscErrorCode initialize() noexcept 45a2158755SJunchao Zhang { 46a2158755SJunchao Zhang PetscFunctionBegin; 47a2158755SJunchao Zhang if (devInitialized_) PetscFunctionReturn(0); 48a2158755SJunchao Zhang if (syclDevice_.is_gpu() && use_gpu_aware_mpi) { 49a2158755SJunchao Zhang if (!isMPISyclAware_()) { 50a2158755SJunchao Zhang (*PetscErrorPrintf)("PETSc is configured with sycl support, but your MPI is not aware of sycl GPU devices. For better performance, please use a sycl GPU-aware MPI.\n"); 51a2158755SJunchao Zhang (*PetscErrorPrintf)("If you do not care, add option -use_gpu_aware_mpi 0. To not see the message again, add the option to your .petscrc, OR add it to the env var PETSC_OPTIONS.\n"); 52a2158755SJunchao Zhang PETSCABORT(PETSC_COMM_SELF,PETSC_ERR_LIB); 53a2158755SJunchao Zhang } 54a2158755SJunchao Zhang } 55a2158755SJunchao Zhang devInitialized_ = true; 56a2158755SJunchao Zhang PetscFunctionReturn(0); 57a2158755SJunchao Zhang } 58a2158755SJunchao Zhang 59a2158755SJunchao Zhang PETSC_NODISCARD PetscErrorCode view(PetscViewer viewer) const noexcept 60a2158755SJunchao Zhang { 61a2158755SJunchao Zhang PetscErrorCode ierr; 62a2158755SJunchao Zhang MPI_Comm comm; 63a2158755SJunchao Zhang PetscMPIInt rank; 64a2158755SJunchao Zhang PetscBool iascii; 65a2158755SJunchao Zhang 66a2158755SJunchao Zhang PetscFunctionBegin; 67*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!devInitialized_,PETSC_COMM_SELF,PETSC_ERR_COR,"Device %d being viewed before it was initialized or configured",id_); 68a2158755SJunchao Zhang ierr = PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer),PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 69a2158755SJunchao Zhang ierr = PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer),&comm);CHKERRQ(ierr); 70a2158755SJunchao Zhang if (iascii) { 71a2158755SJunchao Zhang ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 72a2158755SJunchao Zhang ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 73a2158755SJunchao Zhang ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] device: %s\n",rank,syclDevice_.get_info<sycl::info::device::name>().c_str());CHKERRQ(ierr); 74a2158755SJunchao Zhang // flush the assignment information 75a2158755SJunchao Zhang ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 76a2158755SJunchao Zhang ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr); 77a2158755SJunchao Zhang ierr = PetscViewerASCIISynchronizedPrintf(viewer,"-> Device vendor: %s\n",syclDevice_.get_info<sycl::info::device::vendor>().c_str());CHKERRQ(ierr); 78a2158755SJunchao Zhang ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr); 79a2158755SJunchao Zhang ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 80a2158755SJunchao Zhang ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 81a2158755SJunchao Zhang } 82a2158755SJunchao Zhang PetscFunctionReturn(0); 83a2158755SJunchao Zhang } 84a2158755SJunchao Zhang 85a2158755SJunchao Zhang private: 86a2158755SJunchao Zhang static sycl::device chooseSYCLDevice_(int id) 87a2158755SJunchao Zhang { 88a2158755SJunchao Zhang if (id == PETSC_SYCL_DEVICE_HOST) { 89a2158755SJunchao Zhang return sycl::device(sycl::host_selector()); 90a2158755SJunchao Zhang } else { 91a2158755SJunchao Zhang return sycl::device::get_devices(sycl::info::device_type::gpu)[id]; 92a2158755SJunchao Zhang } 93a2158755SJunchao Zhang } 94a2158755SJunchao Zhang 95a2158755SJunchao Zhang // Is the underlying MPI aware of sycl (GPU) devices? 96a2158755SJunchao Zhang bool isMPISyclAware_() noexcept 97a2158755SJunchao Zhang { 98a2158755SJunchao Zhang PetscErrorCode ierr; 99a2158755SJunchao Zhang const int bufSize = 2; 100a2158755SJunchao Zhang const int hbuf[bufSize] = {1,0}; 101a2158755SJunchao Zhang int *dbuf = nullptr; 102a2158755SJunchao Zhang bool awareness = false; 103a2158755SJunchao Zhang const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 104a2158755SJunchao Zhang if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer,1); 105a2158755SJunchao Zhang return PetscSignalHandlerDefault(signal,ptr); 106a2158755SJunchao Zhang }; 107a2158755SJunchao Zhang 108a2158755SJunchao Zhang PetscFunctionBegin; 109a2158755SJunchao Zhang auto Q = sycl::queue(syclDevice_); 110a2158755SJunchao Zhang dbuf = sycl::malloc_device<int>(bufSize,Q); 111a2158755SJunchao Zhang Q.memcpy(dbuf,hbuf,sizeof(int)*bufSize).wait(); 112a2158755SJunchao Zhang ierr = PetscPushSignalHandler(SyclSignalHandler,nullptr);CHKERRABORT(PETSC_COMM_SELF,ierr); 113a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = true; 114a2158755SJunchao Zhang if (setjmp(MPISyclAwareJumpBuffer)) { 115a2158755SJunchao Zhang // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 116a2158755SJunchao Zhang awareness = false; 117a2158755SJunchao Zhang PetscStackPop; 118a2158755SJunchao Zhang } else if (!MPI_Allreduce(dbuf,dbuf+1,1,MPI_INT,MPI_SUM,PETSC_COMM_SELF)) awareness = true; 119a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = false; 120a2158755SJunchao Zhang ierr = PetscPopSignalHandler();CHKERRABORT(PETSC_COMM_SELF,ierr); 121a2158755SJunchao Zhang sycl::free(dbuf,Q); 122a2158755SJunchao Zhang PetscFunctionReturn(awareness); 123a2158755SJunchao Zhang } 124a2158755SJunchao Zhang }; 125a2158755SJunchao Zhang 12617f48955SJacob Faibussowitsch PetscErrorCode Device::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscDeviceInitType *defaultInitType) noexcept 127a2158755SJunchao Zhang { 128a2158755SJunchao Zhang PetscInt initType = *defaultInitType,id = *defaultDeviceId; 129a2158755SJunchao Zhang PetscBool view = PETSC_FALSE,flg; 130a2158755SJunchao Zhang PetscInt ngpus; 131a2158755SJunchao Zhang PetscErrorCode ierr; 132a2158755SJunchao Zhang 133a2158755SJunchao Zhang PetscFunctionBegin; 134a2158755SJunchao Zhang if (initialized_) PetscFunctionReturn(0); 135a2158755SJunchao Zhang initialized_ = true; 136a2158755SJunchao Zhang ierr = PetscRegisterFinalize(finalize_);CHKERRQ(ierr); 137a2158755SJunchao Zhang 138a2158755SJunchao Zhang ierr = PetscOptionsBegin(comm,nullptr,"PetscDevice SYCL Options","Sys");CHKERRQ(ierr); 139a2158755SJunchao Zhang ierr = PetscOptionsEList("-device_enable_sycl","How (or whether) to initialize a device","SyclDevice::initialize()",PetscDeviceInitTypes,3,PetscDeviceInitTypes[initType],&initType,nullptr);CHKERRQ(ierr); 140a2158755SJunchao Zhang ierr = PetscOptionsRangeInt("-device_select_sycl","Which sycl device to use? Pass -2 for host, PETSC_DECIDE (-1) to let PETSc decide, 0 and up for GPUs","PetscDeviceCreate",id,&id,nullptr,-2,std::numeric_limits<decltype(ngpus)>::max());CHKERRQ(ierr); 141a2158755SJunchao Zhang ierr = PetscOptionsBool("-device_view_sycl","Display device information and assignments (forces eager initialization)",nullptr,view,&view,&flg);CHKERRQ(ierr); 142a2158755SJunchao Zhang ierr = PetscOptionsEnd();CHKERRQ(ierr); 143a2158755SJunchao Zhang 144a2158755SJunchao Zhang // post-process the options and lay the groundwork for initialization if needs be 145a2158755SJunchao Zhang std::vector<sycl::device> gpu_devices = sycl::device::get_devices(sycl::info::device_type::gpu); 146a2158755SJunchao Zhang ngpus = static_cast<PetscInt>(gpu_devices.size()); 147*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(ngpus == 0 && id >= 0,comm,PETSC_ERR_USER_INPUT,"You specified a sycl gpu device with -device_select_sycl %d but there is no GPU", (int)id); 148*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(ngpus > 0 && id >= ngpus,comm,PETSC_ERR_USER_INPUT,"You specified a sycl gpu device with -device_select_sycl %d but there are only %d GPU", (int)id, (int)ngpus); 149a2158755SJunchao Zhang 150a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 151a2158755SJunchao Zhang else { 152a2158755SJunchao Zhang ierr = PetscDeviceCheckDeviceCount_Internal(ngpus);CHKERRQ(ierr); 153a2158755SJunchao Zhang if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 154a2158755SJunchao Zhang if (ngpus) { 155a2158755SJunchao Zhang PetscMPIInt rank; 156a2158755SJunchao Zhang ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 157a2158755SJunchao Zhang id = rank % ngpus; 158a2158755SJunchao Zhang } else id = PETSC_SYCL_DEVICE_HOST; 159a2158755SJunchao Zhang } 160a2158755SJunchao Zhang view = static_cast<decltype(view)>(view && flg); 161a2158755SJunchao Zhang if (view) initType = PETSC_DEVICE_INIT_EAGER; 162a2158755SJunchao Zhang } 163a2158755SJunchao Zhang 164a2158755SJunchao Zhang if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 165a2158755SJunchao Zhang 166a2158755SJunchao Zhang defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 167*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(initType == PETSC_DEVICE_INIT_EAGER && id == PETSC_SYCL_DEVICE_NONE,comm,PETSC_ERR_USER_INPUT,"Cannot eagerly initialize sycl devices as you disabled them by -device_enable_sycl none"); 168a2158755SJunchao Zhang 169a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_EAGER) { 17017f48955SJacob Faibussowitsch devices_[defaultDevice_] = new DeviceInternal(defaultDevice_); 171a2158755SJunchao Zhang ierr = devices_[defaultDevice_]->initialize();CHKERRQ(ierr); 172a2158755SJunchao Zhang if (view) { 173a2158755SJunchao Zhang PetscViewer viewer; 174a2158755SJunchao Zhang ierr = PetscLogInitialize();CHKERRQ(ierr); 175a2158755SJunchao Zhang ierr = PetscViewerASCIIGetStdout(comm,&viewer);CHKERRQ(ierr); 176a2158755SJunchao Zhang ierr = devices_[defaultDevice_]->view(viewer);CHKERRQ(ierr); 177a2158755SJunchao Zhang } 178a2158755SJunchao Zhang } 179a2158755SJunchao Zhang 180a2158755SJunchao Zhang // record the results of the initialization 181a2158755SJunchao Zhang *defaultInitType = static_cast<PetscDeviceInitType>(initType); 182a2158755SJunchao Zhang *defaultDeviceId = id; 183a2158755SJunchao Zhang PetscFunctionReturn(0); 184a2158755SJunchao Zhang } 185a2158755SJunchao Zhang 18617f48955SJacob Faibussowitsch PetscErrorCode Device::finalize_() noexcept 187a2158755SJunchao Zhang { 188a2158755SJunchao Zhang PetscFunctionBegin; 189a2158755SJunchao Zhang if (!initialized_) PetscFunctionReturn(0); 190a2158755SJunchao Zhang for (auto&& devPtr : devices_array_) delete devPtr; 191a2158755SJunchao Zhang defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 192a2158755SJunchao Zhang initialized_ = false; 193a2158755SJunchao Zhang PetscFunctionReturn(0); 194a2158755SJunchao Zhang } 195a2158755SJunchao Zhang 19617f48955SJacob Faibussowitsch PetscErrorCode Device::getDevice(PetscDevice device, PetscInt id) const noexcept 197a2158755SJunchao Zhang { 198a2158755SJunchao Zhang PetscErrorCode ierr; 199a2158755SJunchao Zhang 200a2158755SJunchao Zhang PetscFunctionBegin; 201*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(defaultDevice_ == PETSC_SYCL_DEVICE_NONE,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Trying to retrieve a SYCL PetscDevice when it has been disabled"); 202a2158755SJunchao Zhang if (id == PETSC_DECIDE) id = defaultDevice_; 203*2c71b3e2SJacob Faibussowitsch PetscCheckFalse((id < PETSC_SYCL_DEVICE_HOST) || (id-PETSC_SYCL_DEVICE_HOST >= PETSC_DEVICE_MAX_DEVICES),PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only supports %zu number of devices but trying to get device with id %" PetscInt_FMT,devices_array_.size(),id); 204a2158755SJunchao Zhang if (devices_[id]) { 205*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(id != devices_[id]->id(),PETSC_COMM_SELF,PETSC_ERR_PLIB,"Entry %" PetscInt_FMT " contains device with mismatching id %" PetscInt_FMT,id,devices_[id]->id()); 20617f48955SJacob Faibussowitsch } else devices_[id] = new DeviceInternal(id); 207a2158755SJunchao Zhang ierr = devices_[id]->initialize();CHKERRQ(ierr); 208a2158755SJunchao Zhang device->deviceId = devices_[id]->id(); // technically id = devices_[id]->id_ here 209a2158755SJunchao Zhang device->ops->createcontext = create_; 210a2158755SJunchao Zhang device->ops->configure = this->configureDevice; 211a2158755SJunchao Zhang device->ops->view = this->viewDevice; 212a2158755SJunchao Zhang PetscFunctionReturn(0); 213a2158755SJunchao Zhang } 214a2158755SJunchao Zhang 21517f48955SJacob Faibussowitsch PetscErrorCode Device::configureDevice(PetscDevice device) noexcept 216a2158755SJunchao Zhang { 217a2158755SJunchao Zhang PetscFunctionBegin; 218a2158755SJunchao Zhang // Nothing for now 219a2158755SJunchao Zhang PetscFunctionReturn(0); 220a2158755SJunchao Zhang } 221a2158755SJunchao Zhang 22217f48955SJacob Faibussowitsch PetscErrorCode Device::viewDevice(PetscDevice device, PetscViewer viewer) noexcept 223a2158755SJunchao Zhang { 224a2158755SJunchao Zhang PetscErrorCode ierr; 225a2158755SJunchao Zhang 226a2158755SJunchao Zhang PetscFunctionBegin; 227a2158755SJunchao Zhang ierr = devices_[device->deviceId]->view(viewer);CHKERRQ(ierr); 228a2158755SJunchao Zhang PetscFunctionReturn(0); 229a2158755SJunchao Zhang } 230a2158755SJunchao Zhang 23117f48955SJacob Faibussowitsch } // namespace SYCL 23217f48955SJacob Faibussowitsch 23317f48955SJacob Faibussowitsch } // namespace Device 23417f48955SJacob Faibussowitsch 235a2158755SJunchao Zhang } // namespace Petsc 236