1a2158755SJunchao Zhang #include "../../interface/sycldevice.hpp" 2a2158755SJunchao Zhang #include <csetjmp> // for MPI sycl device awareness 3a2158755SJunchao Zhang #include <csignal> // SIGSEGV 4a2158755SJunchao Zhang #include <vector> 5a2158755SJunchao Zhang #include <CL/sycl.hpp> 6a2158755SJunchao Zhang 7a2158755SJunchao Zhang #if PetscDefined(USE_LOG) 8a2158755SJunchao Zhang PETSC_INTERN PetscErrorCode PetscLogInitialize(void); 9a2158755SJunchao Zhang #else 10a2158755SJunchao Zhang #define PetscLogInitialize() 0 11a2158755SJunchao Zhang #endif 12a2158755SJunchao Zhang 139371c9d4SSatish Balay namespace Petsc { 14a2158755SJunchao Zhang 159371c9d4SSatish Balay namespace Device { 1617f48955SJacob Faibussowitsch 179371c9d4SSatish Balay namespace SYCL { 1817f48955SJacob Faibussowitsch 19a2158755SJunchao Zhang // definition for static 2017f48955SJacob Faibussowitsch std::array<Device::DeviceInternal *, PETSC_DEVICE_MAX_DEVICES> Device::devices_array_ = {}; 2117f48955SJacob Faibussowitsch Device::DeviceInternal **Device::devices_ = &Device::devices_array_[1]; 2217f48955SJacob Faibussowitsch int Device::defaultDevice_ = PETSC_SYCL_DEVICE_NONE; 2317f48955SJacob Faibussowitsch bool Device::initialized_ = false; 24a2158755SJunchao Zhang 25a2158755SJunchao Zhang static std::jmp_buf MPISyclAwareJumpBuffer; 26a2158755SJunchao Zhang static bool MPISyclAwareJumpBufferSet; 27a2158755SJunchao Zhang 28a2158755SJunchao Zhang // internal "impls" class for SyclDevice. Each instance represents a single sycl device 299371c9d4SSatish Balay class PETSC_NODISCARD Device::DeviceInternal { 30a2158755SJunchao Zhang const int id_; // -1 for the host device; 0 and up for gpu devices 31a2158755SJunchao Zhang bool devInitialized_; 32a2158755SJunchao Zhang const sycl::device syclDevice_; 33a2158755SJunchao Zhang 34a2158755SJunchao Zhang public: 35a2158755SJunchao Zhang // default constructor 3617f48955SJacob Faibussowitsch DeviceInternal(int id) noexcept : id_(id), devInitialized_(false), syclDevice_(chooseSYCLDevice_(id)) { } 37a2158755SJunchao Zhang int id() const { return id_; } 38a2158755SJunchao Zhang bool initialized() const { return devInitialized_; } 39a2158755SJunchao Zhang 409371c9d4SSatish Balay PETSC_NODISCARD PetscErrorCode initialize() noexcept { 41a2158755SJunchao Zhang PetscFunctionBegin; 42a2158755SJunchao Zhang if (devInitialized_) PetscFunctionReturn(0); 43a2158755SJunchao Zhang if (syclDevice_.is_gpu() && use_gpu_aware_mpi) { 44a2158755SJunchao Zhang if (!isMPISyclAware_()) { 45a2158755SJunchao 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"); 46a2158755SJunchao 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"); 47a2158755SJunchao Zhang PETSCABORT(PETSC_COMM_SELF, PETSC_ERR_LIB); 48a2158755SJunchao Zhang } 49a2158755SJunchao Zhang } 50a2158755SJunchao Zhang devInitialized_ = true; 51a2158755SJunchao Zhang PetscFunctionReturn(0); 52a2158755SJunchao Zhang } 53a2158755SJunchao Zhang 549371c9d4SSatish Balay PETSC_NODISCARD PetscErrorCode view(PetscViewer viewer) const noexcept { 55a2158755SJunchao Zhang MPI_Comm comm; 56a2158755SJunchao Zhang PetscMPIInt rank; 57a2158755SJunchao Zhang PetscBool iascii; 58a2158755SJunchao Zhang 59a2158755SJunchao Zhang PetscFunctionBegin; 6028b400f6SJacob Faibussowitsch PetscCheck(devInitialized_, PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d being viewed before it was initialized or configured", id_); 619566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare(reinterpret_cast<PetscObject>(viewer), PETSCVIEWERASCII, &iascii)); 629566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm(reinterpret_cast<PetscObject>(viewer), &comm)); 63a2158755SJunchao Zhang if (iascii) { 64c0b02e83SStefano Zampini PetscViewer sviewer; 65c0b02e83SStefano Zampini 669566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 679566063dSJacob Faibussowitsch PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer)); 689566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(sviewer, "[%d] device: %s\n", rank, syclDevice_.get_info<sycl::info::device::name>().c_str())); 699566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPushTab(sviewer)); 709566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(sviewer, "-> Device vendor: %s\n", syclDevice_.get_info<sycl::info::device::vendor>().c_str())); 719566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopTab(sviewer)); 729566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(sviewer)); 739566063dSJacob Faibussowitsch PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer)); 749566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(viewer)); 75a2158755SJunchao Zhang } 76a2158755SJunchao Zhang PetscFunctionReturn(0); 77a2158755SJunchao Zhang } 78a2158755SJunchao Zhang 79*a16fd2c9SJacob Faibussowitsch PETSC_NODISCARD PetscErrorCode getattribute(PetscDeviceAttribute attr, void *value) const noexcept { 80*a16fd2c9SJacob Faibussowitsch PetscFunctionBegin; 81*a16fd2c9SJacob Faibussowitsch PetscCheck(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d not initialized", id()); 82*a16fd2c9SJacob Faibussowitsch switch (attr) { 83*a16fd2c9SJacob Faibussowitsch 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>(); 84*a16fd2c9SJacob Faibussowitsch case PETSC_DEVICE_ATTR_MAX: break; 85*a16fd2c9SJacob Faibussowitsch } 86*a16fd2c9SJacob Faibussowitsch PetscFunctionReturn(0); 87*a16fd2c9SJacob Faibussowitsch } 88*a16fd2c9SJacob Faibussowitsch 89a2158755SJunchao Zhang private: 909371c9d4SSatish Balay static sycl::device chooseSYCLDevice_(int id) { 91a2158755SJunchao Zhang if (id == PETSC_SYCL_DEVICE_HOST) { 92a2158755SJunchao Zhang return sycl::device(sycl::host_selector()); 93a2158755SJunchao Zhang } else { 94a2158755SJunchao Zhang return sycl::device::get_devices(sycl::info::device_type::gpu)[id]; 95a2158755SJunchao Zhang } 96a2158755SJunchao Zhang } 97a2158755SJunchao Zhang 98a2158755SJunchao Zhang // Is the underlying MPI aware of sycl (GPU) devices? 999371c9d4SSatish Balay bool isMPISyclAware_() noexcept { 100a2158755SJunchao Zhang const int bufSize = 2; 101a2158755SJunchao Zhang const int hbuf[bufSize] = {1, 0}; 102a2158755SJunchao Zhang int *dbuf = nullptr; 103a2158755SJunchao Zhang bool awareness = false; 104a2158755SJunchao Zhang const auto SyclSignalHandler = [](int signal, void *ptr) -> PetscErrorCode { 105a2158755SJunchao Zhang if ((signal == SIGSEGV) && MPISyclAwareJumpBufferSet) std::longjmp(MPISyclAwareJumpBuffer, 1); 106a2158755SJunchao Zhang return PetscSignalHandlerDefault(signal, ptr); 107a2158755SJunchao Zhang }; 108a2158755SJunchao Zhang 109a2158755SJunchao Zhang PetscFunctionBegin; 110a2158755SJunchao Zhang auto Q = sycl::queue(syclDevice_); 111a2158755SJunchao Zhang dbuf = sycl::malloc_device<int>(bufSize, Q); 112a2158755SJunchao Zhang Q.memcpy(dbuf, hbuf, sizeof(int) * bufSize).wait(); 1139566063dSJacob Faibussowitsch PetscCallAbort(PETSC_COMM_SELF, PetscPushSignalHandler(SyclSignalHandler, nullptr)); 114a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = true; 115a2158755SJunchao Zhang if (setjmp(MPISyclAwareJumpBuffer)) { 116a2158755SJunchao Zhang // if a segv was triggered in the MPI_Allreduce below, it is very likely due to MPI not being GPU-aware 117a2158755SJunchao Zhang awareness = false; 118a2158755SJunchao Zhang PetscStackPop; 119a2158755SJunchao Zhang } else if (!MPI_Allreduce(dbuf, dbuf + 1, 1, MPI_INT, MPI_SUM, PETSC_COMM_SELF)) awareness = true; 120a2158755SJunchao Zhang MPISyclAwareJumpBufferSet = false; 1219566063dSJacob Faibussowitsch PetscCallAbort(PETSC_COMM_SELF, PetscPopSignalHandler()); 122a2158755SJunchao Zhang sycl::free(dbuf, Q); 123a2158755SJunchao Zhang PetscFunctionReturn(awareness); 124a2158755SJunchao Zhang } 125a2158755SJunchao Zhang }; 126a2158755SJunchao Zhang 1279371c9d4SSatish Balay PetscErrorCode Device::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscDeviceInitType *defaultInitType) noexcept { 128a2158755SJunchao Zhang PetscInt initType = *defaultInitType, id = *defaultDeviceId; 129a2158755SJunchao Zhang PetscBool view = PETSC_FALSE, flg; 130a2158755SJunchao Zhang PetscInt ngpus; 131a2158755SJunchao Zhang 132a2158755SJunchao Zhang PetscFunctionBegin; 133a2158755SJunchao Zhang if (initialized_) PetscFunctionReturn(0); 134a2158755SJunchao Zhang initialized_ = true; 1359566063dSJacob Faibussowitsch PetscCall(PetscRegisterFinalize(finalize_)); 136a2158755SJunchao Zhang 137d0609cedSBarry Smith PetscOptionsBegin(comm, nullptr, "PetscDevice SYCL Options", "Sys"); 1389566063dSJacob Faibussowitsch PetscCall(PetscOptionsEList("-device_enable_sycl", "How (or whether) to initialize a device", "SyclDevice::initialize()", PetscDeviceInitTypes, 3, PetscDeviceInitTypes[initType], &initType, nullptr)); 1399566063dSJacob Faibussowitsch 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())); 1409566063dSJacob Faibussowitsch PetscCall(PetscOptionsBool("-device_view_sycl", "Display device information and assignments (forces eager initialization)", nullptr, view, &view, &flg)); 141d0609cedSBarry Smith PetscOptionsEnd(); 142a2158755SJunchao Zhang 143a2158755SJunchao Zhang // post-process the options and lay the groundwork for initialization if needs be 144a2158755SJunchao Zhang std::vector<sycl::device> gpu_devices = sycl::device::get_devices(sycl::info::device_type::gpu); 145a2158755SJunchao Zhang ngpus = static_cast<PetscInt>(gpu_devices.size()); 146cc73adaaSBarry Smith 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); 147cc73adaaSBarry Smith 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); 148a2158755SJunchao Zhang 149a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_NONE) id = PETSC_SYCL_DEVICE_NONE; /* user wants to disable all sycl devices */ 150a2158755SJunchao Zhang else { 1519566063dSJacob Faibussowitsch PetscCall(PetscDeviceCheckDeviceCount_Internal(ngpus)); 152a2158755SJunchao Zhang if (id == PETSC_DECIDE) { /* petsc will choose a GPU device if any, otherwise a CPU device */ 153a2158755SJunchao Zhang if (ngpus) { 154a2158755SJunchao Zhang PetscMPIInt rank; 1559566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 156a2158755SJunchao Zhang id = rank % ngpus; 157a2158755SJunchao Zhang } else id = PETSC_SYCL_DEVICE_HOST; 158a2158755SJunchao Zhang } 159a2158755SJunchao Zhang view = static_cast<decltype(view)>(view && flg); 160a2158755SJunchao Zhang if (view) initType = PETSC_DEVICE_INIT_EAGER; 161a2158755SJunchao Zhang } 162a2158755SJunchao Zhang 163a2158755SJunchao Zhang if (id == -2) id = PETSC_SYCL_DEVICE_HOST; // user passed in '-device_select_sycl -2'. We transform it into canonical form 164a2158755SJunchao Zhang 165a2158755SJunchao Zhang defaultDevice_ = static_cast<decltype(defaultDevice_)>(id); 166cc73adaaSBarry Smith 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"); 167a2158755SJunchao Zhang 168a2158755SJunchao Zhang if (initType == PETSC_DEVICE_INIT_EAGER) { 16917f48955SJacob Faibussowitsch devices_[defaultDevice_] = new DeviceInternal(defaultDevice_); 1709566063dSJacob Faibussowitsch PetscCall(devices_[defaultDevice_]->initialize()); 171a2158755SJunchao Zhang if (view) { 172a2158755SJunchao Zhang PetscViewer viewer; 1739566063dSJacob Faibussowitsch PetscCall(PetscLogInitialize()); 1749566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIGetStdout(comm, &viewer)); 1759566063dSJacob Faibussowitsch PetscCall(devices_[defaultDevice_]->view(viewer)); 176a2158755SJunchao Zhang } 177a2158755SJunchao Zhang } 178a2158755SJunchao Zhang 179a2158755SJunchao Zhang // record the results of the initialization 180a2158755SJunchao Zhang *defaultInitType = static_cast<PetscDeviceInitType>(initType); 181a2158755SJunchao Zhang *defaultDeviceId = id; 182a2158755SJunchao Zhang PetscFunctionReturn(0); 183a2158755SJunchao Zhang } 184a2158755SJunchao Zhang 1859371c9d4SSatish Balay PetscErrorCode Device::finalize_() noexcept { 186a2158755SJunchao Zhang PetscFunctionBegin; 187a2158755SJunchao Zhang if (!initialized_) PetscFunctionReturn(0); 188a2158755SJunchao Zhang for (auto &&devPtr : devices_array_) delete devPtr; 189a2158755SJunchao Zhang defaultDevice_ = PETSC_SYCL_DEVICE_NONE; // disabled by default 190a2158755SJunchao Zhang initialized_ = false; 191a2158755SJunchao Zhang PetscFunctionReturn(0); 192a2158755SJunchao Zhang } 193a2158755SJunchao Zhang 1949371c9d4SSatish Balay PetscErrorCode Device::getDevice(PetscDevice device, PetscInt id) const noexcept { 195a2158755SJunchao Zhang PetscFunctionBegin; 19608401ef6SPierre Jolivet PetscCheck(defaultDevice_ != PETSC_SYCL_DEVICE_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Trying to retrieve a SYCL PetscDevice when it has been disabled"); 197a2158755SJunchao Zhang if (id == PETSC_DECIDE) id = defaultDevice_; 1989371c9d4SSatish Balay 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, 1999371c9d4SSatish Balay devices_array_.size(), id); 200a2158755SJunchao Zhang if (devices_[id]) { 20108401ef6SPierre Jolivet 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()); 20217f48955SJacob Faibussowitsch } else devices_[id] = new DeviceInternal(id); 2039566063dSJacob Faibussowitsch PetscCall(devices_[id]->initialize()); 204a2158755SJunchao Zhang device->deviceId = devices_[id]->id(); // technically id = devices_[id]->id_ here 205a2158755SJunchao Zhang device->ops->createcontext = create_; 206a2158755SJunchao Zhang device->ops->configure = this->configureDevice; 207a2158755SJunchao Zhang device->ops->view = this->viewDevice; 208*a16fd2c9SJacob Faibussowitsch device->ops->getattribute = this->getAttribute; 209a2158755SJunchao Zhang PetscFunctionReturn(0); 210a2158755SJunchao Zhang } 211a2158755SJunchao Zhang 2129371c9d4SSatish Balay PetscErrorCode Device::configureDevice(PetscDevice device) noexcept { 213a2158755SJunchao Zhang PetscFunctionBegin; 214a2158755SJunchao Zhang // Nothing for now 215a2158755SJunchao Zhang PetscFunctionReturn(0); 216a2158755SJunchao Zhang } 217a2158755SJunchao Zhang 2189371c9d4SSatish Balay PetscErrorCode Device::viewDevice(PetscDevice device, PetscViewer viewer) noexcept { 219a2158755SJunchao Zhang PetscFunctionBegin; 2209566063dSJacob Faibussowitsch PetscCall(devices_[device->deviceId]->view(viewer)); 221a2158755SJunchao Zhang PetscFunctionReturn(0); 222a2158755SJunchao Zhang } 223a2158755SJunchao Zhang 224*a16fd2c9SJacob Faibussowitsch PetscErrorCode Device::getAttribute(PetscDevice device, PetscDeviceAttribute attr, void *value) noexcept { 225*a16fd2c9SJacob Faibussowitsch PetscFunctionBegin; 226*a16fd2c9SJacob Faibussowitsch PetscCall(devices_[device->deviceId]->getattribute(attr, value)); 227*a16fd2c9SJacob Faibussowitsch PetscFunctionReturn(0); 228*a16fd2c9SJacob Faibussowitsch } 229*a16fd2c9SJacob Faibussowitsch 23017f48955SJacob Faibussowitsch } // namespace SYCL 23117f48955SJacob Faibussowitsch 23217f48955SJacob Faibussowitsch } // namespace Device 23317f48955SJacob Faibussowitsch 234a2158755SJunchao Zhang } // namespace Petsc 235