1 #include "sycldevice.hpp" 2 #include <limits> // for std::numeric_limits 3 #include <csetjmp> // for MPI sycl device awareness 4 #include <csignal> // SIGSEGV 5 #include <vector> 6 #include <CL/sycl.hpp> 7 8 namespace Petsc { 9 10 namespace device { 11 12 namespace sycl { 13 14 // definition for static 15 std::array<Device::DeviceInternal *, PETSC_DEVICE_MAX_DEVICES> Device::devices_array_ = {}; 16 Device::DeviceInternal **Device::devices_ = &Device::devices_array_[1]; 17 int Device::defaultDevice_ = PETSC_SYCL_DEVICE_NONE; 18 bool Device::initialized_ = false; 19 20 static std::jmp_buf MPISyclAwareJumpBuffer; 21 static bool MPISyclAwareJumpBufferSet; 22 23 // internal "impls" class for SyclDevice. Each instance represents a single sycl device 24 class PETSC_NODISCARD Device::DeviceInternal { 25 const int id_; // -1 for the host device; 0 and up for gpu devices 26 bool devInitialized_; 27 const ::sycl::device syclDevice_; 28 29 public: 30 // default constructor 31 DeviceInternal(int id) noexcept : id_(id), devInitialized_(false), syclDevice_(chooseSYCLDevice_(id)) { } 32 int id() const { return id_; } 33 bool initialized() const { return devInitialized_; } 34 35 PETSC_NODISCARD PetscErrorCode initialize() noexcept { 36 PetscFunctionBegin; 37 if (initialized()) PetscFunctionReturn(0); 38 if (syclDevice_.is_gpu() && use_gpu_aware_mpi) { 39 if (!isMPISyclAware_()) { 40 (*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"); 41 (*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"); 42 PETSCABORT(PETSC_COMM_SELF, PETSC_ERR_LIB); 43 } 44 } 45 devInitialized_ = true; 46 PetscFunctionReturn(0); 47 } 48 49 PETSC_NODISCARD PetscErrorCode view(PetscViewer viewer) const noexcept { 50 MPI_Comm comm; 51 PetscMPIInt rank; 52 PetscBool iascii; 53 54 PetscFunctionBegin; 55 PetscCheck(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d being viewed before it was initialized or configured", id()); 56 PetscCall(PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer), PETSCVIEWERASCII, &iascii)); 57 PetscCall(PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer), &comm)); 58 if (iascii) { 59 PetscViewer sviewer; 60 61 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 62 PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer)); 63 PetscCall(PetscViewerASCIIPrintf(sviewer, "[%d] device: %s\n", rank, syclDevice_.get_info<::sycl::info::device::name>().c_str())); 64 PetscCall(PetscViewerASCIIPushTab(sviewer)); 65 PetscCall(PetscViewerASCIIPrintf(sviewer, "-> Device vendor: %s\n", syclDevice_.get_info<::sycl::info::device::vendor>().c_str())); 66 PetscCall(PetscViewerASCIIPopTab(sviewer)); 67 PetscCall(PetscViewerFlush(sviewer)); 68 PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer)); 69 PetscCall(PetscViewerFlush(viewer)); 70 } 71 PetscFunctionReturn(0); 72 } 73 74 PETSC_NODISCARD PetscErrorCode getattribute(PetscDeviceAttribute attr, void *value) const noexcept { 75 PetscFunctionBegin; 76 PetscCheck(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d not initialized", id()); 77 switch (attr) { 78 case PETSC_DEVICE_ATTR_SIZE_T_SHARED_MEM_PER_BLOCK: *static_cast<std::size_t *>(value) = syclDevice_.get_info<::sycl::info::device::local_mem_size>(); 79 case PETSC_DEVICE_ATTR_MAX: break; 80 } 81 PetscFunctionReturn(0); 82 } 83 84 private: 85 static ::sycl::device chooseSYCLDevice_(int id) { 86 if (id == PETSC_SYCL_DEVICE_HOST) { 87 return ::sycl::device(::sycl::host_selector()); 88 } else { 89 return ::sycl::device::get_devices(::sycl::info::device_type::gpu)[id]; 90 } 91 } 92 93 // Is the underlying MPI aware of sycl (GPU) devices? 94 bool isMPISyclAware_() noexcept { 95 const int bufSize = 2; 96 const int hbuf[bufSize] = {1, 0}; 97 int *dbuf = nullptr; 98 bool awareness = false; 99 const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 100 if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer, 1); 101 return PetscSignalHandlerDefault(signal, ptr); 102 }; 103 104 PetscFunctionBegin; 105 auto Q = ::sycl::queue(syclDevice_); 106 dbuf = ::sycl::malloc_device<int>(bufSize, Q); 107 Q.memcpy(dbuf, hbuf, sizeof(int) * bufSize).wait(); 108 PetscCallAbort(PETSC_COMM_SELF, PetscPushSignalHandler(SyclSignalHandler, nullptr)); 109 MPISyclAwareJumpBufferSet = true; 110 if (setjmp(MPISyclAwareJumpBuffer)) { 111 // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 112 awareness = false; 113 PetscStackPop; 114 } else if (!MPI_Allreduce(dbuf, dbuf + 1, 1, MPI_INT, MPI_SUM, PETSC_COMM_SELF)) awareness = true; 115 MPISyclAwareJumpBufferSet = false; 116 PetscCallAbort(PETSC_COMM_SELF, PetscPopSignalHandler()); 117 ::sycl::free(dbuf, Q); 118 PetscFunctionReturn(awareness); 119 } 120 }; 121 122 PetscErrorCode Device::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscBool *defaultView, PetscDeviceInitType *defaultInitType) noexcept { 123 auto id = *defaultDeviceId; 124 auto initType = *defaultInitType; 125 auto view = *defaultView, flg = PETSC_FALSE; 126 PetscInt ngpus; 127 128 PetscFunctionBegin; 129 if (initialized_) PetscFunctionReturn(0); 130 initialized_ = true; 131 PetscCall(PetscRegisterFinalize(finalize_)); 132 PetscOptionsBegin(comm, nullptr, "PetscDevice sycl Options", "Sys"); 133 PetscCall(base_type::PetscOptionDeviceInitialize(PetscOptionsObject, &initType, nullptr)); 134 PetscCall(base_type::PetscOptionDeviceSelect(PetscOptionsObject, "Which sycl device to use? Pass -2 for host, PETSC_DECIDE (" PetscStringize(PETSC_DECIDE) ") to let PETSc decide, 0 and up for GPUs", "PetscDeviceCreate()", id, &id, nullptr, -2, std::numeric_limits<decltype(ngpus)>::max())); 135 static_assert(PETSC_DECIDE - 1 == -2, ""); 136 PetscCall(base_type::PetscOptionDeviceView(PetscOptionsObject, &view, &flg)); 137 PetscOptionsEnd(); 138 139 // post-process the options and lay the groundwork for initialization if needs be 140 std::vector<::sycl::device> gpu_devices = ::sycl::device::get_devices(::sycl::info::device_type::gpu); 141 ngpus = static_cast<PetscInt>(gpu_devices.size()); 142 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); 143 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); 144 145 if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 146 else { 147 PetscCall(PetscDeviceCheckDeviceCount_Internal(ngpus)); 148 if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 149 if (ngpus) { 150 PetscMPIInt rank; 151 PetscCallMPI(MPI_Comm_rank(comm, &rank)); 152 id = rank % ngpus; 153 } else id = PETSC_SYCL_DEVICE_HOST; 154 } 155 if (view) initType = PETSC_DEVICE_INIT_EAGER; 156 } 157 158 if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 159 160 defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 161 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"); 162 // record the results of the initialization 163 *defaultDeviceId = id; 164 *defaultView = view; 165 *defaultInitType = initType; 166 PetscFunctionReturn(0); 167 } 168 169 PetscErrorCode Device::finalize_() noexcept { 170 PetscFunctionBegin; 171 if (!initialized_) PetscFunctionReturn(0); 172 for (auto &&devPtr : devices_array_) delete devPtr; 173 defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 174 initialized_ = false; 175 PetscFunctionReturn(0); 176 } 177 178 PetscErrorCode Device::init_device_id_(PetscInt *inid) const noexcept { 179 const auto id = *inid == PETSC_DECIDE ? defaultDevice_ : *inid; 180 181 PetscFunctionBegin; 182 PetscCheck(defaultDevice_ != PETSC_SYCL_DEVICE_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Trying to retrieve a SYCL PetscDevice when it has been disabled"); 183 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, 184 devices_array_.size(), id); 185 if (!devices_[id]) devices_[id] = new DeviceInternal(id); 186 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()); 187 PetscCall(devices_[id]->initialize()); 188 *inid = id; 189 PetscFunctionReturn(0); 190 } 191 192 PetscErrorCode Device::view_device_(PetscDevice device, PetscViewer viewer) noexcept { 193 PetscFunctionBegin; 194 PetscCall(devices_[device->deviceId]->view(viewer)); 195 PetscFunctionReturn(0); 196 } 197 198 PetscErrorCode Device::get_attribute_(PetscInt id, PetscDeviceAttribute attr, void *value) noexcept { 199 PetscFunctionBegin; 200 PetscCall(devices_[id]->getattribute(attr, value)); 201 PetscFunctionReturn(0); 202 } 203 204 } // namespace sycl 205 206 } // namespace device 207 208 } // namespace Petsc 209