1 #include "../../interface/sycldevice.hpp" 2 #include <csetjmp> // for MPI sycl device awareness 3 #include <csignal> // SIGSEGV 4 #include <vector> 5 #include <CL/sycl.hpp> 6 7 #if PetscDefined(USE_LOG) 8 PETSC_INTERN PetscErrorCode PetscLogInitialize(void); 9 #else 10 #define PetscLogInitialize() 0 11 #endif 12 13 namespace Petsc 14 { 15 16 namespace Device 17 { 18 19 namespace SYCL 20 { 21 22 // definition for static 23 std::array<Device::DeviceInternal*,PETSC_DEVICE_MAX_DEVICES> Device::devices_array_ = {}; 24 Device::DeviceInternal** Device::devices_ = &Device::devices_array_[1]; 25 int Device::defaultDevice_ = PETSC_SYCL_DEVICE_NONE; 26 bool Device::initialized_ = false; 27 28 static std::jmp_buf MPISyclAwareJumpBuffer; 29 static bool MPISyclAwareJumpBufferSet; 30 31 // internal "impls" class for SyclDevice. Each instance represents a single sycl device 32 class PETSC_NODISCARD Device::DeviceInternal 33 { 34 const int id_; // -1 for the host device; 0 and up for gpu devices 35 bool devInitialized_; 36 const sycl::device syclDevice_; 37 38 public: 39 // default constructor 40 DeviceInternal(int id) noexcept : id_(id),devInitialized_(false),syclDevice_(chooseSYCLDevice_(id)){} 41 int id() const {return id_;} 42 bool initialized() const {return devInitialized_;} 43 44 PETSC_NODISCARD PetscErrorCode initialize() noexcept 45 { 46 PetscFunctionBegin; 47 if (devInitialized_) PetscFunctionReturn(0); 48 if (syclDevice_.is_gpu() && use_gpu_aware_mpi) { 49 if (!isMPISyclAware_()) { 50 (*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"); 51 (*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"); 52 PETSCABORT(PETSC_COMM_SELF,PETSC_ERR_LIB); 53 } 54 } 55 devInitialized_ = true; 56 PetscFunctionReturn(0); 57 } 58 59 PETSC_NODISCARD PetscErrorCode view(PetscViewer viewer) const noexcept 60 { 61 PetscErrorCode ierr; 62 MPI_Comm comm; 63 PetscMPIInt rank; 64 PetscBool iascii; 65 66 PetscFunctionBegin; 67 if (PetscUnlikely(!devInitialized_)) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_COR,"Device %d being viewed before it was initialized or configured",id_); 68 ierr = PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer),PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 69 ierr = PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer),&comm);CHKERRQ(ierr); 70 if (iascii) { 71 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 72 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 73 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] device: %s\n",rank,syclDevice_.get_info<sycl::info::device::name>().c_str());CHKERRQ(ierr); 74 // flush the assignment information 75 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 76 ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr); 77 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"-> Device vendor: %s\n",syclDevice_.get_info<sycl::info::device::vendor>().c_str());CHKERRQ(ierr); 78 ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr); 79 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 80 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 81 } 82 PetscFunctionReturn(0); 83 } 84 85 private: 86 static sycl::device chooseSYCLDevice_(int id) 87 { 88 if (id == PETSC_SYCL_DEVICE_HOST) { 89 return sycl::device(sycl::host_selector()); 90 } else { 91 return sycl::device::get_devices(sycl::info::device_type::gpu)[id]; 92 } 93 } 94 95 // Is the underlying MPI aware of sycl (GPU) devices? 96 bool isMPISyclAware_() noexcept 97 { 98 PetscErrorCode ierr; 99 const int bufSize = 2; 100 const int hbuf[bufSize] = {1,0}; 101 int *dbuf = nullptr; 102 bool awareness = false; 103 const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 104 if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer,1); 105 return PetscSignalHandlerDefault(signal,ptr); 106 }; 107 108 PetscFunctionBegin; 109 auto Q = sycl::queue(syclDevice_); 110 dbuf = sycl::malloc_device<int>(bufSize,Q); 111 Q.memcpy(dbuf,hbuf,sizeof(int)*bufSize).wait(); 112 ierr = PetscPushSignalHandler(SyclSignalHandler,nullptr);CHKERRABORT(PETSC_COMM_SELF,ierr); 113 MPISyclAwareJumpBufferSet = true; 114 if (setjmp(MPISyclAwareJumpBuffer)) { 115 // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 116 awareness = false; 117 PetscStackPop; 118 } else if (!MPI_Allreduce(dbuf,dbuf+1,1,MPI_INT,MPI_SUM,PETSC_COMM_SELF)) awareness = true; 119 MPISyclAwareJumpBufferSet = false; 120 ierr = PetscPopSignalHandler();CHKERRABORT(PETSC_COMM_SELF,ierr); 121 sycl::free(dbuf,Q); 122 PetscFunctionReturn(awareness); 123 } 124 }; 125 126 PetscErrorCode Device::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscDeviceInitType *defaultInitType) noexcept 127 { 128 PetscInt initType = *defaultInitType,id = *defaultDeviceId; 129 PetscBool view = PETSC_FALSE,flg; 130 PetscInt ngpus; 131 PetscErrorCode ierr; 132 133 PetscFunctionBegin; 134 if (initialized_) PetscFunctionReturn(0); 135 initialized_ = true; 136 ierr = PetscRegisterFinalize(finalize_);CHKERRQ(ierr); 137 138 ierr = PetscOptionsBegin(comm,nullptr,"PetscDevice SYCL Options","Sys");CHKERRQ(ierr); 139 ierr = PetscOptionsEList("-device_enable_sycl","How (or whether) to initialize a device","SyclDevice::initialize()",PetscDeviceInitTypes,3,PetscDeviceInitTypes[initType],&initType,nullptr);CHKERRQ(ierr); 140 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); 141 ierr = PetscOptionsBool("-device_view_sycl","Display device information and assignments (forces eager initialization)",nullptr,view,&view,&flg);CHKERRQ(ierr); 142 ierr = PetscOptionsEnd();CHKERRQ(ierr); 143 144 // post-process the options and lay the groundwork for initialization if needs be 145 std::vector<sycl::device> gpu_devices = sycl::device::get_devices(sycl::info::device_type::gpu); 146 ngpus = static_cast<PetscInt>(gpu_devices.size()); 147 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); 148 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); 149 150 if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 151 else { 152 ierr = PetscDeviceCheckDeviceCount_Internal(ngpus);CHKERRQ(ierr); 153 if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 154 if (ngpus) { 155 PetscMPIInt rank; 156 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 157 id = rank % ngpus; 158 } else id = PETSC_SYCL_DEVICE_HOST; 159 } 160 view = static_cast<decltype(view)>(view && flg); 161 if (view) initType = PETSC_DEVICE_INIT_EAGER; 162 } 163 164 if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 165 166 defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 167 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"); 168 169 if (initType == PETSC_DEVICE_INIT_EAGER) { 170 devices_[defaultDevice_] = new DeviceInternal(defaultDevice_); 171 ierr = devices_[defaultDevice_]->initialize();CHKERRQ(ierr); 172 if (view) { 173 PetscViewer viewer; 174 ierr = PetscLogInitialize();CHKERRQ(ierr); 175 ierr = PetscViewerASCIIGetStdout(comm,&viewer);CHKERRQ(ierr); 176 ierr = devices_[defaultDevice_]->view(viewer);CHKERRQ(ierr); 177 } 178 } 179 180 // record the results of the initialization 181 *defaultInitType = static_cast<PetscDeviceInitType>(initType); 182 *defaultDeviceId = id; 183 PetscFunctionReturn(0); 184 } 185 186 PetscErrorCode Device::finalize_() noexcept 187 { 188 PetscFunctionBegin; 189 if (!initialized_) PetscFunctionReturn(0); 190 for (auto&& devPtr : devices_array_) delete devPtr; 191 defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 192 initialized_ = false; 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode Device::getDevice(PetscDevice device, PetscInt id) const noexcept 197 { 198 PetscErrorCode ierr; 199 200 PetscFunctionBegin; 201 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"); 202 if (id == PETSC_DECIDE) id = defaultDevice_; 203 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); 204 if (devices_[id]) { 205 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()); 206 } else devices_[id] = new DeviceInternal(id); 207 ierr = devices_[id]->initialize();CHKERRQ(ierr); 208 device->deviceId = devices_[id]->id(); // technically id = devices_[id]->id_ here 209 device->ops->createcontext = create_; 210 device->ops->configure = this->configureDevice; 211 device->ops->view = this->viewDevice; 212 PetscFunctionReturn(0); 213 } 214 215 PetscErrorCode Device::configureDevice(PetscDevice device) noexcept 216 { 217 PetscFunctionBegin; 218 // Nothing for now 219 PetscFunctionReturn(0); 220 } 221 222 PetscErrorCode Device::viewDevice(PetscDevice device, PetscViewer viewer) noexcept 223 { 224 PetscErrorCode ierr; 225 226 PetscFunctionBegin; 227 ierr = devices_[device->deviceId]->view(viewer);CHKERRQ(ierr); 228 PetscFunctionReturn(0); 229 } 230 231 } // namespace SYCL 232 233 } // namespace Device 234 235 } // namespace Petsc 236