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