1*a2158755SJunchao Zhang #include "../../interface/sycldevice.hpp" 2*a2158755SJunchao Zhang #include <csetjmp> // for MPI sycl device awareness 3*a2158755SJunchao Zhang #include <csignal> // SIGSEGV 4*a2158755SJunchao Zhang #include <vector> 5*a2158755SJunchao Zhang #include <CL/sycl.hpp> 6*a2158755SJunchao Zhang 7*a2158755SJunchao Zhang #if PetscDefined(USE_LOG) 8*a2158755SJunchao Zhang PETSC_INTERN PetscErrorCode PetscLogInitialize(void); 9*a2158755SJunchao Zhang #else 10*a2158755SJunchao Zhang #define PetscLogInitialize() 0 11*a2158755SJunchao Zhang #endif 12*a2158755SJunchao Zhang 13*a2158755SJunchao Zhang namespace Petsc 14*a2158755SJunchao Zhang { 15*a2158755SJunchao Zhang 16*a2158755SJunchao Zhang // definition for static 17*a2158755SJunchao Zhang std::array<SyclDevice::SyclDeviceInternal*,PETSC_DEVICE_MAX_DEVICES> SyclDevice::devices_array_ = {}; 18*a2158755SJunchao Zhang SyclDevice::SyclDeviceInternal** SyclDevice::devices_ = &SyclDevice::devices_array_[1]; 19*a2158755SJunchao Zhang int SyclDevice::defaultDevice_ = PETSC_SYCL_DEVICE_NONE; 20*a2158755SJunchao Zhang bool SyclDevice::initialized_ = false; 21*a2158755SJunchao Zhang 22*a2158755SJunchao Zhang static std::jmp_buf MPISyclAwareJumpBuffer; 23*a2158755SJunchao Zhang static bool MPISyclAwareJumpBufferSet; 24*a2158755SJunchao Zhang 25*a2158755SJunchao Zhang // internal "impls" class for SyclDevice. Each instance represents a single sycl device 26*a2158755SJunchao Zhang class PETSC_NODISCARD SyclDevice::SyclDeviceInternal 27*a2158755SJunchao Zhang { 28*a2158755SJunchao Zhang const int id_; // -1 for the host device; 0 and up for gpu devices 29*a2158755SJunchao Zhang bool devInitialized_; 30*a2158755SJunchao Zhang const sycl::device syclDevice_; 31*a2158755SJunchao Zhang 32*a2158755SJunchao Zhang public: 33*a2158755SJunchao Zhang // default constructor 34*a2158755SJunchao Zhang SyclDeviceInternal(int id) noexcept : id_(id),devInitialized_(false),syclDevice_(chooseSYCLDevice_(id)){} 35*a2158755SJunchao Zhang int id() const {return id_;} 36*a2158755SJunchao Zhang bool initialized() const {return devInitialized_;} 37*a2158755SJunchao Zhang 38*a2158755SJunchao Zhang PETSC_NODISCARD PetscErrorCode initialize() noexcept 39*a2158755SJunchao Zhang { 40*a2158755SJunchao Zhang PetscFunctionBegin; 41*a2158755SJunchao Zhang if (devInitialized_) PetscFunctionReturn(0); 42*a2158755SJunchao Zhang if (syclDevice_.is_gpu() && use_gpu_aware_mpi) { 43*a2158755SJunchao Zhang if (!isMPISyclAware_()) { 44*a2158755SJunchao 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"); 45*a2158755SJunchao 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"); 46*a2158755SJunchao Zhang PETSCABORT(PETSC_COMM_SELF,PETSC_ERR_LIB); 47*a2158755SJunchao Zhang } 48*a2158755SJunchao Zhang } 49*a2158755SJunchao Zhang devInitialized_ = true; 50*a2158755SJunchao Zhang PetscFunctionReturn(0); 51*a2158755SJunchao Zhang } 52*a2158755SJunchao Zhang 53*a2158755SJunchao Zhang PETSC_NODISCARD PetscErrorCode view(PetscViewer viewer) const noexcept 54*a2158755SJunchao Zhang { 55*a2158755SJunchao Zhang PetscErrorCode ierr; 56*a2158755SJunchao Zhang MPI_Comm comm; 57*a2158755SJunchao Zhang PetscMPIInt rank; 58*a2158755SJunchao Zhang PetscBool iascii; 59*a2158755SJunchao Zhang 60*a2158755SJunchao Zhang PetscFunctionBegin; 61*a2158755SJunchao Zhang if (PetscUnlikely(!devInitialized_)) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_COR,"Device %d being viewed before it was initialized or configured",id_); 62*a2158755SJunchao Zhang ierr = PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer),PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 63*a2158755SJunchao Zhang ierr = PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer),&comm);CHKERRQ(ierr); 64*a2158755SJunchao Zhang if (iascii) { 65*a2158755SJunchao Zhang ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 66*a2158755SJunchao Zhang ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 67*a2158755SJunchao Zhang ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] device: %s\n",rank,syclDevice_.get_info<sycl::info::device::name>().c_str());CHKERRQ(ierr); 68*a2158755SJunchao Zhang // flush the assignment information 69*a2158755SJunchao Zhang ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 70*a2158755SJunchao Zhang ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr); 71*a2158755SJunchao Zhang ierr = PetscViewerASCIISynchronizedPrintf(viewer,"-> Device vendor: %s\n",syclDevice_.get_info<sycl::info::device::vendor>().c_str());CHKERRQ(ierr); 72*a2158755SJunchao Zhang ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr); 73*a2158755SJunchao Zhang ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 74*a2158755SJunchao Zhang ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 75*a2158755SJunchao Zhang } 76*a2158755SJunchao Zhang PetscFunctionReturn(0); 77*a2158755SJunchao Zhang } 78*a2158755SJunchao Zhang 79*a2158755SJunchao Zhang private: 80*a2158755SJunchao Zhang static sycl::device chooseSYCLDevice_(int id) 81*a2158755SJunchao Zhang { 82*a2158755SJunchao Zhang if (id == PETSC_SYCL_DEVICE_HOST) { 83*a2158755SJunchao Zhang return sycl::device(sycl::host_selector()); 84*a2158755SJunchao Zhang } else { 85*a2158755SJunchao Zhang return sycl::device::get_devices(sycl::info::device_type::gpu)[id]; 86*a2158755SJunchao Zhang } 87*a2158755SJunchao Zhang } 88*a2158755SJunchao Zhang 89*a2158755SJunchao Zhang // Is the underlying MPI aware of sycl (GPU) devices? 90*a2158755SJunchao Zhang bool isMPISyclAware_() noexcept 91*a2158755SJunchao Zhang { 92*a2158755SJunchao Zhang PetscErrorCode ierr; 93*a2158755SJunchao Zhang const int bufSize = 2; 94*a2158755SJunchao Zhang const int hbuf[bufSize] = {1,0}; 95*a2158755SJunchao Zhang int *dbuf = nullptr; 96*a2158755SJunchao Zhang bool awareness = false; 97*a2158755SJunchao Zhang const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 98*a2158755SJunchao Zhang if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer,1); 99*a2158755SJunchao Zhang return PetscSignalHandlerDefault(signal,ptr); 100*a2158755SJunchao Zhang }; 101*a2158755SJunchao Zhang 102*a2158755SJunchao Zhang PetscFunctionBegin; 103*a2158755SJunchao Zhang auto Q = sycl::queue(syclDevice_); 104*a2158755SJunchao Zhang dbuf = sycl::malloc_device<int>(bufSize,Q); 105*a2158755SJunchao Zhang Q.memcpy(dbuf,hbuf,sizeof(int)*bufSize).wait(); 106*a2158755SJunchao Zhang ierr = PetscPushSignalHandler(SyclSignalHandler,nullptr);CHKERRABORT(PETSC_COMM_SELF,ierr); 107*a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = true; 108*a2158755SJunchao Zhang if (setjmp(MPISyclAwareJumpBuffer)) { 109*a2158755SJunchao Zhang // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 110*a2158755SJunchao Zhang awareness = false; 111*a2158755SJunchao Zhang PetscStackPop; 112*a2158755SJunchao Zhang } else if (!MPI_Allreduce(dbuf,dbuf+1,1,MPI_INT,MPI_SUM,PETSC_COMM_SELF)) awareness = true; 113*a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = false; 114*a2158755SJunchao Zhang ierr = PetscPopSignalHandler();CHKERRABORT(PETSC_COMM_SELF,ierr); 115*a2158755SJunchao Zhang sycl::free(dbuf,Q); 116*a2158755SJunchao Zhang PetscFunctionReturn(awareness); 117*a2158755SJunchao Zhang } 118*a2158755SJunchao Zhang }; 119*a2158755SJunchao Zhang 120*a2158755SJunchao Zhang PetscErrorCode SyclDevice::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscDeviceInitType *defaultInitType) noexcept 121*a2158755SJunchao Zhang { 122*a2158755SJunchao Zhang PetscInt initType = *defaultInitType,id = *defaultDeviceId; 123*a2158755SJunchao Zhang PetscBool view = PETSC_FALSE,flg; 124*a2158755SJunchao Zhang PetscInt ngpus; 125*a2158755SJunchao Zhang PetscErrorCode ierr; 126*a2158755SJunchao Zhang 127*a2158755SJunchao Zhang PetscFunctionBegin; 128*a2158755SJunchao Zhang if (initialized_) PetscFunctionReturn(0); 129*a2158755SJunchao Zhang initialized_ = true; 130*a2158755SJunchao Zhang ierr = PetscRegisterFinalize(finalize_);CHKERRQ(ierr); 131*a2158755SJunchao Zhang 132*a2158755SJunchao Zhang ierr = PetscOptionsBegin(comm,nullptr,"PetscDevice SYCL Options","Sys");CHKERRQ(ierr); 133*a2158755SJunchao Zhang ierr = PetscOptionsEList("-device_enable_sycl","How (or whether) to initialize a device","SyclDevice::initialize()",PetscDeviceInitTypes,3,PetscDeviceInitTypes[initType],&initType,nullptr);CHKERRQ(ierr); 134*a2158755SJunchao 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); 135*a2158755SJunchao Zhang ierr = PetscOptionsBool("-device_view_sycl","Display device information and assignments (forces eager initialization)",nullptr,view,&view,&flg);CHKERRQ(ierr); 136*a2158755SJunchao Zhang ierr = PetscOptionsEnd();CHKERRQ(ierr); 137*a2158755SJunchao Zhang 138*a2158755SJunchao Zhang // post-process the options and lay the groundwork for initialization if needs be 139*a2158755SJunchao Zhang std::vector<sycl::device> gpu_devices = sycl::device::get_devices(sycl::info::device_type::gpu); 140*a2158755SJunchao Zhang ngpus = static_cast<PetscInt>(gpu_devices.size()); 141*a2158755SJunchao Zhang if (PetscUnlikely(ngpus == 0 && id >= 0)) SETERRQ1(comm,PETSC_ERR_USER_INPUT,"You specified a sycl gpu device with -device_select_sycl %d but there is no GPU", (int)id); 142*a2158755SJunchao Zhang if (PetscUnlikely(ngpus > 0 && id >= ngpus)) SETERRQ2(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); 143*a2158755SJunchao Zhang 144*a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 145*a2158755SJunchao Zhang else { 146*a2158755SJunchao Zhang ierr = PetscDeviceCheckDeviceCount_Internal(ngpus);CHKERRQ(ierr); 147*a2158755SJunchao Zhang if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 148*a2158755SJunchao Zhang if (ngpus) { 149*a2158755SJunchao Zhang PetscMPIInt rank; 150*a2158755SJunchao Zhang ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 151*a2158755SJunchao Zhang id = rank % ngpus; 152*a2158755SJunchao Zhang } else id = PETSC_SYCL_DEVICE_HOST; 153*a2158755SJunchao Zhang } 154*a2158755SJunchao Zhang view = static_cast<decltype(view)>(view && flg); 155*a2158755SJunchao Zhang if (view) initType = PETSC_DEVICE_INIT_EAGER; 156*a2158755SJunchao Zhang } 157*a2158755SJunchao Zhang 158*a2158755SJunchao Zhang if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 159*a2158755SJunchao Zhang 160*a2158755SJunchao Zhang defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 161*a2158755SJunchao Zhang if (PetscUnlikely(initType == PETSC_DEVICE_INIT_EAGER && id == PETSC_SYCL_DEVICE_NONE)) SETERRQ(comm,PETSC_ERR_USER_INPUT,"Cannot eagerly initialize sycl devices as you disabled them by -device_enable_sycl none"); 162*a2158755SJunchao Zhang 163*a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_EAGER) { 164*a2158755SJunchao Zhang devices_[defaultDevice_] = new SyclDeviceInternal(defaultDevice_); 165*a2158755SJunchao Zhang ierr = devices_[defaultDevice_]->initialize();CHKERRQ(ierr); 166*a2158755SJunchao Zhang if (view) { 167*a2158755SJunchao Zhang PetscViewer viewer; 168*a2158755SJunchao Zhang ierr = PetscLogInitialize();CHKERRQ(ierr); 169*a2158755SJunchao Zhang ierr = PetscViewerASCIIGetStdout(comm,&viewer);CHKERRQ(ierr); 170*a2158755SJunchao Zhang ierr = devices_[defaultDevice_]->view(viewer);CHKERRQ(ierr); 171*a2158755SJunchao Zhang } 172*a2158755SJunchao Zhang } 173*a2158755SJunchao Zhang 174*a2158755SJunchao Zhang // record the results of the initialization 175*a2158755SJunchao Zhang *defaultInitType = static_cast<PetscDeviceInitType>(initType); 176*a2158755SJunchao Zhang *defaultDeviceId = id; 177*a2158755SJunchao Zhang PetscFunctionReturn(0); 178*a2158755SJunchao Zhang } 179*a2158755SJunchao Zhang 180*a2158755SJunchao Zhang PetscErrorCode SyclDevice::finalize_() noexcept 181*a2158755SJunchao Zhang { 182*a2158755SJunchao Zhang PetscFunctionBegin; 183*a2158755SJunchao Zhang if (!initialized_) PetscFunctionReturn(0); 184*a2158755SJunchao Zhang for (auto&& devPtr : devices_array_) delete devPtr; 185*a2158755SJunchao Zhang defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 186*a2158755SJunchao Zhang initialized_ = false; 187*a2158755SJunchao Zhang PetscFunctionReturn(0); 188*a2158755SJunchao Zhang } 189*a2158755SJunchao Zhang 190*a2158755SJunchao Zhang PetscErrorCode SyclDevice::getDevice(PetscDevice device, PetscInt id) const noexcept 191*a2158755SJunchao Zhang { 192*a2158755SJunchao Zhang PetscErrorCode ierr; 193*a2158755SJunchao Zhang 194*a2158755SJunchao Zhang PetscFunctionBegin; 195*a2158755SJunchao Zhang if (PetscUnlikely(defaultDevice_ == PETSC_SYCL_DEVICE_NONE)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Trying to retrieve a SYCL PetscDevice when it has been disabled"); 196*a2158755SJunchao Zhang if (id == PETSC_DECIDE) id = defaultDevice_; 197*a2158755SJunchao Zhang if ((id < PETSC_SYCL_DEVICE_HOST) || (id-PETSC_SYCL_DEVICE_HOST >= PETSC_DEVICE_MAX_DEVICES)) SETERRQ2(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); 198*a2158755SJunchao Zhang if (devices_[id]) { 199*a2158755SJunchao Zhang if (id != devices_[id]->id()) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Entry %" PetscInt_FMT " contains device with mismatching id %" PetscInt_FMT,id,devices_[id]->id()); 200*a2158755SJunchao Zhang } else devices_[id] = new SyclDeviceInternal(id); 201*a2158755SJunchao Zhang ierr = devices_[id]->initialize();CHKERRQ(ierr); 202*a2158755SJunchao Zhang device->deviceId = devices_[id]->id(); // technically id = devices_[id]->id_ here 203*a2158755SJunchao Zhang device->ops->createcontext = create_; 204*a2158755SJunchao Zhang device->ops->configure = this->configureDevice; 205*a2158755SJunchao Zhang device->ops->view = this->viewDevice; 206*a2158755SJunchao Zhang PetscFunctionReturn(0); 207*a2158755SJunchao Zhang } 208*a2158755SJunchao Zhang 209*a2158755SJunchao Zhang PetscErrorCode SyclDevice::configureDevice(PetscDevice device) noexcept 210*a2158755SJunchao Zhang { 211*a2158755SJunchao Zhang PetscFunctionBegin; 212*a2158755SJunchao Zhang // Nothing for now 213*a2158755SJunchao Zhang PetscFunctionReturn(0); 214*a2158755SJunchao Zhang } 215*a2158755SJunchao Zhang 216*a2158755SJunchao Zhang PetscErrorCode SyclDevice::viewDevice(PetscDevice device, PetscViewer viewer) noexcept 217*a2158755SJunchao Zhang { 218*a2158755SJunchao Zhang PetscErrorCode ierr; 219*a2158755SJunchao Zhang 220*a2158755SJunchao Zhang PetscFunctionBegin; 221*a2158755SJunchao Zhang ierr = devices_[device->deviceId]->view(viewer);CHKERRQ(ierr); 222*a2158755SJunchao Zhang PetscFunctionReturn(0); 223*a2158755SJunchao Zhang } 224*a2158755SJunchao Zhang 225*a2158755SJunchao Zhang } // namespace Petsc 226