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 MPI_Comm comm; 62 PetscMPIInt rank; 63 PetscBool iascii; 64 65 PetscFunctionBegin; 66 PetscCheck(devInitialized_,PETSC_COMM_SELF,PETSC_ERR_COR,"Device %d being viewed before it was initialized or configured",id_); 67 PetscCall(PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer),PETSCVIEWERASCII,&iascii)); 68 PetscCall(PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer),&comm)); 69 if (iascii) { 70 PetscViewer sviewer; 71 72 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 73 PetscCall(PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer)); 74 PetscCall(PetscViewerASCIIPrintf(sviewer,"[%d] device: %s\n",rank,syclDevice_.get_info<sycl::info::device::name>().c_str())); 75 PetscCall(PetscViewerASCIIPushTab(sviewer)); 76 PetscCall(PetscViewerASCIIPrintf(sviewer,"-> Device vendor: %s\n",syclDevice_.get_info<sycl::info::device::vendor>().c_str())); 77 PetscCall(PetscViewerASCIIPopTab(sviewer)); 78 PetscCall(PetscViewerFlush(sviewer)); 79 PetscCall(PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer)); 80 PetscCall(PetscViewerFlush(viewer)); 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 const int bufSize = 2; 99 const int hbuf[bufSize] = {1,0}; 100 int *dbuf = nullptr; 101 bool awareness = false; 102 const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 103 if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer,1); 104 return PetscSignalHandlerDefault(signal,ptr); 105 }; 106 107 PetscFunctionBegin; 108 auto Q = sycl::queue(syclDevice_); 109 dbuf = sycl::malloc_device<int>(bufSize,Q); 110 Q.memcpy(dbuf,hbuf,sizeof(int)*bufSize).wait(); 111 PetscCallAbort(PETSC_COMM_SELF,PetscPushSignalHandler(SyclSignalHandler,nullptr)); 112 MPISyclAwareJumpBufferSet = true; 113 if (setjmp(MPISyclAwareJumpBuffer)) { 114 // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 115 awareness = false; 116 PetscStackPop; 117 } else if (!MPI_Allreduce(dbuf,dbuf+1,1,MPI_INT,MPI_SUM,PETSC_COMM_SELF)) awareness = true; 118 MPISyclAwareJumpBufferSet = false; 119 PetscCallAbort(PETSC_COMM_SELF,PetscPopSignalHandler()); 120 sycl::free(dbuf,Q); 121 PetscFunctionReturn(awareness); 122 } 123 }; 124 125 PetscErrorCode Device::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscDeviceInitType *defaultInitType) noexcept 126 { 127 PetscInt initType = *defaultInitType,id = *defaultDeviceId; 128 PetscBool view = PETSC_FALSE,flg; 129 PetscInt ngpus; 130 131 PetscFunctionBegin; 132 if (initialized_) PetscFunctionReturn(0); 133 initialized_ = true; 134 PetscCall(PetscRegisterFinalize(finalize_)); 135 136 PetscOptionsBegin(comm,nullptr,"PetscDevice SYCL Options","Sys"); 137 PetscCall(PetscOptionsEList("-device_enable_sycl","How (or whether) to initialize a device","SyclDevice::initialize()",PetscDeviceInitTypes,3,PetscDeviceInitTypes[initType],&initType,nullptr)); 138 PetscCall(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())); 139 PetscCall(PetscOptionsBool("-device_view_sycl","Display device information and assignments (forces eager initialization)",nullptr,view,&view,&flg)); 140 PetscOptionsEnd(); 141 142 // post-process the options and lay the groundwork for initialization if needs be 143 std::vector<sycl::device> gpu_devices = sycl::device::get_devices(sycl::info::device_type::gpu); 144 ngpus = static_cast<PetscInt>(gpu_devices.size()); 145 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); 146 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); 147 148 if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 149 else { 150 PetscCall(PetscDeviceCheckDeviceCount_Internal(ngpus)); 151 if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 152 if (ngpus) { 153 PetscMPIInt rank; 154 PetscCallMPI(MPI_Comm_rank(comm,&rank)); 155 id = rank % ngpus; 156 } else id = PETSC_SYCL_DEVICE_HOST; 157 } 158 view = static_cast<decltype(view)>(view && flg); 159 if (view) initType = PETSC_DEVICE_INIT_EAGER; 160 } 161 162 if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 163 164 defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 165 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"); 166 167 if (initType == PETSC_DEVICE_INIT_EAGER) { 168 devices_[defaultDevice_] = new DeviceInternal(defaultDevice_); 169 PetscCall(devices_[defaultDevice_]->initialize()); 170 if (view) { 171 PetscViewer viewer; 172 PetscCall(PetscLogInitialize()); 173 PetscCall(PetscViewerASCIIGetStdout(comm,&viewer)); 174 PetscCall(devices_[defaultDevice_]->view(viewer)); 175 } 176 } 177 178 // record the results of the initialization 179 *defaultInitType = static_cast<PetscDeviceInitType>(initType); 180 *defaultDeviceId = id; 181 PetscFunctionReturn(0); 182 } 183 184 PetscErrorCode Device::finalize_() noexcept 185 { 186 PetscFunctionBegin; 187 if (!initialized_) PetscFunctionReturn(0); 188 for (auto&& devPtr : devices_array_) delete devPtr; 189 defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 190 initialized_ = false; 191 PetscFunctionReturn(0); 192 } 193 194 PetscErrorCode Device::getDevice(PetscDevice device, PetscInt id) const noexcept 195 { 196 PetscFunctionBegin; 197 PetscCheck(defaultDevice_ != PETSC_SYCL_DEVICE_NONE,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Trying to retrieve a SYCL PetscDevice when it has been disabled"); 198 if (id == PETSC_DECIDE) id = defaultDevice_; 199 PetscCheck(!(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); 200 if (devices_[id]) { 201 PetscCheck(id == devices_[id]->id(),PETSC_COMM_SELF,PETSC_ERR_PLIB,"Entry %" PetscInt_FMT " contains device with mismatching id %" PetscInt_FMT,id,devices_[id]->id()); 202 } else devices_[id] = new DeviceInternal(id); 203 PetscCall(devices_[id]->initialize()); 204 device->deviceId = devices_[id]->id(); // technically id = devices_[id]->id_ here 205 device->ops->createcontext = create_; 206 device->ops->configure = this->configureDevice; 207 device->ops->view = this->viewDevice; 208 PetscFunctionReturn(0); 209 } 210 211 PetscErrorCode Device::configureDevice(PetscDevice device) noexcept 212 { 213 PetscFunctionBegin; 214 // Nothing for now 215 PetscFunctionReturn(0); 216 } 217 218 PetscErrorCode Device::viewDevice(PetscDevice device, PetscViewer viewer) noexcept 219 { 220 PetscFunctionBegin; 221 PetscCall(devices_[device->deviceId]->view(viewer)); 222 PetscFunctionReturn(0); 223 } 224 225 } // namespace SYCL 226 227 } // namespace Device 228 229 } // namespace Petsc 230