Actual source code: cupmdevice.cxx

  1: #include <petsc/private/cpp/memory.hpp>

  3: #include "cupmdevice.hpp"

  5: #include <algorithm>
  6: #include <csetjmp> // for cuda mpi awareness
  7: #include <csignal> // SIGSEGV
  8: #include <iterator>
  9: #include <type_traits>

 11: namespace Petsc
 12: {

 14: namespace device
 15: {

 17: namespace cupm
 18: {

 20: // internal "impls" class for CUPMDevice. Each instance represents a single cupm device
 21: template <DeviceType T>
 22: class Device<T>::DeviceInternal {
 23:   const int        id_;
 24:   bool             devInitialized_ = false;
 25:   cupmDeviceProp_t dprop_{}; // cudaDeviceProp appears to be an actual struct, i.e. you can't
 26:                              // initialize it with nullptr or NULL (i've tried)

 28:   static PetscErrorCode CUPMAwareMPI_(bool *) noexcept;

 30: public:
 31:   // default constructor
 32:   explicit constexpr DeviceInternal(int dev) noexcept : id_(dev) { }

 34:   // gather all relevant information for a particular device, a cupmDeviceProp_t is
 35:   // usually sufficient here
 36:   PetscErrorCode initialize() noexcept;
 37:   PetscErrorCode configure() noexcept;
 38:   PetscErrorCode view(PetscViewer) const noexcept;
 39:   PetscErrorCode getattribute(PetscDeviceAttribute, void *) const noexcept;
 40:   PetscErrorCode shutdown() noexcept;

 42:   PETSC_NODISCARD auto id() const -> decltype(id_) { return id_; }
 43:   PETSC_NODISCARD auto initialized() const -> decltype(devInitialized_) { return devInitialized_; }
 44:   PETSC_NODISCARD auto prop() const -> const decltype(dprop_) & { return dprop_; }
 45: };

 47: // the goal here is simply to get the cupm backend to create its context, not to do any type of
 48: // modification of it, or create objects (since these may be affected by subsequent
 49: // configuration changes)
 50: template <DeviceType T>
 51: PetscErrorCode Device<T>::DeviceInternal::initialize() noexcept
 52: {
 53:   PetscFunctionBegin;
 54:   if (initialized()) PetscFunctionReturn(PETSC_SUCCESS);
 55:   devInitialized_ = true;
 56:   // need to do this BEFORE device has been set, although if the user
 57:   // has already done this then we just ignore it
 58:   if (cupmSetDeviceFlags(cupmDeviceMapHost) == cupmErrorSetOnActiveProcess) {
 59:     // reset the error if it was cupmErrorSetOnActiveProcess
 60:     const auto PETSC_UNUSED unused = cupmGetLastError();
 61:   } else PetscCallCUPM(cupmGetLastError());
 62:   // cuda 5.0+ will create a context when cupmSetDevice is called
 63:   if (cupmSetDevice(id()) != cupmErrorDeviceAlreadyInUse) PetscCallCUPM(cupmGetLastError());
 64:   // and in case it doesn't, explicitly call init here
 65:   PetscCallCUPM(cupmInit(0));
 66: #if PetscDefined(HAVE_CUDA)
 67:   // nvmlInit() deprecated in NVML 5.319
 68:   PetscCallNVML(nvmlInit_v2());
 69: #endif
 70:   // where is this variable defined and when is it set? who knows! but it is defined and set
 71:   // at this point. either way, each device must make this check since I guess MPI might not be
 72:   // aware of all of them?
 73:   if (use_gpu_aware_mpi) {
 74:     bool aware;

 76:     // Even the MPI implementation is configured with GPU-aware, it might still need extra settings to enable it.
 77:     // So we do the check at runtime with a code that works only with GPU-aware MPI.
 78:     PetscCall(CUPMAwareMPI_(&aware));
 79:     if (PetscUnlikely(!aware)) {
 80:       PetscCall((*PetscErrorPrintf)("PETSc is configured with GPU support, but your MPI is not GPU-aware. For better performance, please use a GPU-aware MPI.\n"));
 81:       PetscCall((*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"));
 82:       PetscCall((*PetscErrorPrintf)("For Open MPI, you need to configure it with CUDA, ROCm or GPU-aware UCX (https://docs.open-mpi.org/en/main/tuning-apps/accelerators/index.html)\n"));
 83:       PetscCall((*PetscErrorPrintf)("If you already configured it with GPU-aware UCX, you may need 'mpiexec -n <np> --mca pml ucx' or export 'OMPI_MCA_pml=\"ucx\"' to use it.\n"));
 84:       PetscCall((*PetscErrorPrintf)("For MVAPICH2-GDR, you need to set MV2_USE_CUDA=1 (http://mvapich.cse.ohio-state.edu/userguide/gdr/)\n"));
 85:       PetscCall((*PetscErrorPrintf)("For Cray-MPICH, export MPICH_GPU_SUPPORT_ENABLED=1 (see its 'man mpi'); for MPICH, export MPIR_CVAR_ENABLE_GPU=1\n"));
 86:       PETSCABORT(PETSC_COMM_SELF, PETSC_ERR_LIB);
 87:     }
 88:   }
 89:   PetscFunctionReturn(PETSC_SUCCESS);
 90: }

 92: template <DeviceType T>
 93: PetscErrorCode Device<T>::DeviceInternal::configure() noexcept
 94: {
 95:   PetscFunctionBegin;
 96:   PetscAssert(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d being configured before it was initialized", id());
 97:   // why on EARTH nvidia insists on making otherwise informational states into
 98:   // fully-fledged error codes is beyond me. Why couldn't a pointer to bool argument have
 99:   // sufficed?!?!?!
100:   if (cupmSetDevice(id_) != cupmErrorDeviceAlreadyInUse) PetscCallCUPM(cupmGetLastError());
101:   // need to update the device properties
102:   PetscCallCUPM(cupmGetDeviceProperties(&dprop_, id_));
103:   PetscDeviceCUPMRuntimeArch = dprop_.major * 10 + dprop_.minor;
104:   PetscCall(PetscInfo(nullptr, "Configured device %d\n", id_));
105:   PetscFunctionReturn(PETSC_SUCCESS);
106: }

108: template <DeviceType T>
109: PetscErrorCode Device<T>::DeviceInternal::view(PetscViewer viewer) const noexcept
110: {
111:   PetscBool isascii;

113:   PetscFunctionBegin;
114:   PetscAssert(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d being viewed before it was initialized or configured", id());
115:   // we don't print device-specific info in CI-mode
116:   if (PetscUnlikely(PetscCIEnabled)) PetscFunctionReturn(PETSC_SUCCESS);
117:   PetscCall(PetscObjectTypeCompare(PetscObjectCast(viewer), PETSCVIEWERASCII, &isascii));
118:   if (isascii) {
119:     MPI_Comm    comm;
120:     PetscMPIInt rank;
121:     PetscViewer sviewer;

123:     int clock, memclock;
124:     PetscCallCUPM(cupmDeviceGetAttribute(&clock, cupmDevAttrClockRate, id_));
125:     PetscCallCUPM(cupmDeviceGetAttribute(&memclock, cupmDevAttrMemoryClockRate, id_));

127:     PetscCall(PetscObjectGetComm(PetscObjectCast(viewer), &comm));
128:     PetscCallMPI(MPI_Comm_rank(comm, &rank));
129:     PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
130:     PetscCall(PetscViewerASCIIPrintf(sviewer, "[%d] name: %s\n", rank, dprop_.name));
131:     PetscCall(PetscViewerASCIIPushTab(sviewer));
132:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Compute capability: %d.%d\n", dprop_.major, dprop_.minor));
133:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Multiprocessor Count: %d\n", dprop_.multiProcessorCount));
134:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Maximum Grid Dimensions: %d x %d x %d\n", dprop_.maxGridSize[0], dprop_.maxGridSize[1], dprop_.maxGridSize[2]));
135:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Maximum Block Dimensions: %d x %d x %d\n", dprop_.maxThreadsDim[0], dprop_.maxThreadsDim[1], dprop_.maxThreadsDim[2]));
136:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Maximum Threads Per Block: %d\n", dprop_.maxThreadsPerBlock));
137:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Warp Size: %d\n", dprop_.warpSize));
138:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Total Global Memory (bytes): %zu\n", dprop_.totalGlobalMem));
139:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Total Constant Memory (bytes): %zu\n", dprop_.totalConstMem));
140:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Shared Memory Per Block (bytes): %zu\n", dprop_.sharedMemPerBlock));
141:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Multiprocessor Clock Rate (kHz): %d\n", clock));
142:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Memory Clock Rate (kHz): %d\n", memclock));
143:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Memory Bus Width (bits): %d\n", dprop_.memoryBusWidth));
144:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Peak Memory Bandwidth (GB/s): %f\n", 2.0 * memclock * (dprop_.memoryBusWidth / 8) / 1.0e6));
145:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Can map host memory: %s\n", dprop_.canMapHostMemory ? "PETSC_TRUE" : "PETSC_FALSE"));
146:     PetscCall(PetscViewerASCIIPrintf(sviewer, "Can execute multiple kernels concurrently: %s\n", dprop_.concurrentKernels ? "PETSC_TRUE" : "PETSC_FALSE"));
147:     PetscCall(PetscViewerASCIIPopTab(sviewer));
148:     PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
149:   }
150:   PetscFunctionReturn(PETSC_SUCCESS);
151: }

153: template <DeviceType T>
154: PetscErrorCode Device<T>::DeviceInternal::getattribute(PetscDeviceAttribute attr, void *value) const noexcept
155: {
156:   PetscFunctionBegin;
157:   PetscAssert(initialized(), PETSC_COMM_SELF, PETSC_ERR_COR, "Device %d was not initialized", id());
158:   switch (attr) {
159:   case PETSC_DEVICE_ATTR_SIZE_T_SHARED_MEM_PER_BLOCK:
160:     *static_cast<std::size_t *>(value) = prop().sharedMemPerBlock;
161:   case PETSC_DEVICE_ATTR_MAX:
162:     break;
163:   }
164:   PetscFunctionReturn(PETSC_SUCCESS);
165: }

167: template <DeviceType T>
168: PetscErrorCode Device<T>::DeviceInternal::shutdown() noexcept
169: {
170:   PetscFunctionBegin;
171:   if (!initialized()) PetscFunctionReturn(PETSC_SUCCESS);
172: #if PetscDefined(HAVE_CUDA)
173:   PetscCallNVML(nvmlShutdown());
174: #endif
175:   PetscFunctionReturn(PETSC_SUCCESS);
176: }

178: static std::jmp_buf cupmMPIAwareJumpBuffer;
179: static bool         cupmMPIAwareJumpBufferSet;

181: // godspeed to anyone that attempts to call this function
182: void SilenceVariableIsNotNeededAndWillNotBeEmittedWarning_ThisFunctionShouldNeverBeCalled()
183: {
184:   PETSCABORT(MPI_COMM_NULL, (PetscErrorCode)INT_MAX);
185:   if (cupmMPIAwareJumpBufferSet) (void)cupmMPIAwareJumpBuffer;
186: }

188: template <DeviceType T>
189: PetscErrorCode Device<T>::DeviceInternal::CUPMAwareMPI_(bool *awareness) noexcept
190: {
191:   constexpr int hbuf[]            = {1, 0};
192:   int          *dbuf              = nullptr;
193:   const auto    cupmSignalHandler = [](int signal, void *ptr) -> PetscErrorCode {
194:     if ((signal == SIGSEGV) && cupmMPIAwareJumpBufferSet) std::longjmp(cupmMPIAwareJumpBuffer, 1);
195:     return PetscSignalHandlerDefault(signal, ptr);
196:   };

198:   PetscFunctionBegin;
199:   *awareness = false;
200:   PetscCallCUPM(cupmMalloc(reinterpret_cast<void **>(&dbuf), sizeof(hbuf)));
201:   PetscCallCUPM(cupmMemcpy(dbuf, hbuf, sizeof(hbuf), cupmMemcpyHostToDevice));
202:   PetscCallCUPM(cupmDeviceSynchronize());
203:   PetscCall(PetscPushSignalHandler(cupmSignalHandler, nullptr));
204:   cupmMPIAwareJumpBufferSet = true;
205:   if (!setjmp(cupmMPIAwareJumpBuffer) && !MPI_Allreduce(dbuf, dbuf + 1, 1, MPI_INT, MPI_SUM, PETSC_COMM_SELF)) *awareness = true;
206:   cupmMPIAwareJumpBufferSet = false;
207:   PetscCall(PetscPopSignalHandler());
208:   PetscCallCUPM(cupmFree(dbuf));
209:   PetscFunctionReturn(PETSC_SUCCESS);
210: }

212: template <DeviceType T>
213: PetscErrorCode Device<T>::finalize_() noexcept
214: {
215:   PetscFunctionBegin;
216:   if (PetscUnlikely(!initialized_)) PetscFunctionReturn(PETSC_SUCCESS);
217:   for (auto &&device : devices_) {
218:     if (device) PetscCall(device->shutdown());
219:     device.reset();
220:   }
221:   defaultDevice_ = PETSC_CUPM_DEVICE_NONE; // disabled by default
222:   initialized_   = false;
223:   PetscFunctionReturn(PETSC_SUCCESS);
224: }

226: template <DeviceType T>
227: PETSC_NODISCARD static PETSC_CONSTEXPR_14 const char *CUPM_VISIBLE_DEVICES() noexcept
228: {
229:   switch (T) {
230:   case DeviceType::CUDA:
231:     return "CUDA_VISIBLE_DEVICES";
232:   case DeviceType::HIP:
233:     return "HIP_VISIBLE_DEVICES";
234:   }
235:   PetscUnreachable();
236:   return "PETSC_ERROR_PLIB";
237: }

239: /*
240:      The default device ID is
241:        MPI     -- rank % number_local_devices
242:        PyTorch -- getenv("LOCAL_RANK")
243: */
244: template <DeviceType T>
245: PetscErrorCode Device<T>::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscBool *defaultView, PetscDeviceInitType *defaultInitType) noexcept
246: {
247:   auto initId   = std::make_pair(*defaultDeviceId, PETSC_FALSE);
248:   auto initView = std::make_pair(*defaultView, PETSC_FALSE);
249:   auto initType = std::make_pair(*defaultInitType, PETSC_FALSE);
250:   int  ndev     = 0;

252:   PetscFunctionBegin;
253:   if (initialized_) PetscFunctionReturn(PETSC_SUCCESS);
254:   initialized_ = true;
255:   PetscCall(PetscRegisterFinalize(finalize_));
256:   PetscCall(base_type::PetscOptionDeviceAll(comm, initType, initId, initView));

258:   if (initType.first == PETSC_DEVICE_INIT_NONE) {
259:     initId.first = PETSC_CUPM_DEVICE_NONE;
260:   } else if (const auto cerr = cupmGetDeviceCount(&ndev)) {
261:     auto PETSC_UNUSED ignored = cupmGetLastError();

263:     PetscCheck((initType.first != PETSC_DEVICE_INIT_EAGER) && !initView.first, comm, PETSC_ERR_USER_INPUT, "Cannot eagerly initialize %s, as doing so results in %s error %d (%s) : %s", cupmName(), cupmName(), static_cast<PetscErrorCode>(cerr), cupmGetErrorName(cerr), cupmGetErrorString(cerr));
264:     // we won't be initializing anything anyways
265:     initType.first = PETSC_DEVICE_INIT_NONE;
266:     // save the error code for later
267:     initId.first = -static_cast<decltype(initId.first)>(cerr);
268:   }

270:   // check again for init type, since the device count may have changed it
271:   if (initType.first == PETSC_DEVICE_INIT_NONE) {
272:     // id < 0 (excluding PETSC_DECIDE) indicates an error has occurred during setup
273:     if ((initId.first > 0) || (initId.first == PETSC_DECIDE)) initId.first = PETSC_CUPM_DEVICE_NONE;
274:     // initType overrides initView
275:     initView.first = PETSC_FALSE;
276:   } else {
277:     PetscCall(PetscDeviceCheckDeviceCount_Internal(ndev));
278:     if (initId.first == PETSC_DECIDE) {
279:       if (ndev) {
280:         /* TORCHELASTIC_RUN_ID is used as a proxy to determine if the current process was launched with torchrun */
281:         char *pytorch_exists = (char *)getenv("TORCHELASTIC_RUN_ID");
282:         char *pytorch_rank   = (char *)getenv("LOCAL_RANK");

284:         if (pytorch_exists && pytorch_rank) {
285:           char *endptr;

287:           initId.first = (PetscInt)strtol(pytorch_rank, &endptr, 10);
288:           PetscCheck(initId.first < ndev, PETSC_COMM_SELF, PETSC_ERR_LIB, "PyTorch environmental variable LOCAL_RANK %s > number devices %d", pytorch_rank, ndev);
289:         } else {
290:           PetscMPIInt rank;

292:           PetscCallMPI(MPI_Comm_rank(comm, &rank));
293:           initId.first = rank % ndev;
294:         }
295:       } else initId.first = 0;
296:     }
297:     if (initView.first) initType.first = PETSC_DEVICE_INIT_EAGER;
298:   }

300:   static_assert(std::is_same<PetscMPIInt, decltype(defaultDevice_)>::value, "");
301:   // initId.first is PetscInt, _defaultDevice is int
302:   PetscCall(PetscMPIIntCast(initId.first, &defaultDevice_));
303:   // record the results of the initialization
304:   *defaultDeviceId = initId.first;
305:   *defaultView     = initView.first;
306:   *defaultInitType = initType.first;
307:   PetscFunctionReturn(PETSC_SUCCESS);
308: }

310: template <DeviceType T>
311: PetscErrorCode Device<T>::init_device_id_(PetscInt *inid) const noexcept
312: {
313:   const auto id   = *inid == PETSC_DECIDE ? defaultDevice_ : (int)*inid;
314:   const auto cerr = static_cast<cupmError_t>(-defaultDevice_);

316:   PetscFunctionBegin;
317:   PetscCheck(defaultDevice_ != PETSC_CUPM_DEVICE_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Trying to retrieve a %s PetscDevice when it has been disabled", cupmName());
318:   PetscCheck(defaultDevice_ >= 0, PETSC_COMM_SELF, PETSC_ERR_GPU, "Cannot lazily initialize PetscDevice: %s error %d (%s) : %s", cupmName(), static_cast<PetscErrorCode>(cerr), cupmGetErrorName(cerr), cupmGetErrorString(cerr));
319:   PetscAssert(static_cast<decltype(devices_.size())>(id) < devices_.size(), PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only supports %zu number of devices but trying to get device with id %d", devices_.size(), id);

321:   if (!devices_[id]) devices_[id] = util::make_unique<DeviceInternal>(id);
322:   PetscAssert(id == devices_[id]->id(), PETSC_COMM_SELF, PETSC_ERR_PLIB, "Entry %d contains device with mismatching id %d", id, devices_[id]->id());
323:   PetscCall(devices_[id]->initialize());
324:   *inid = id;
325:   PetscFunctionReturn(PETSC_SUCCESS);
326: }

328: template <DeviceType T>
329: PetscErrorCode Device<T>::configure_device_(PetscDevice device) noexcept
330: {
331:   PetscFunctionBegin;
332:   PetscCall(devices_[device->deviceId]->configure());
333:   PetscFunctionReturn(PETSC_SUCCESS);
334: }

336: template <DeviceType T>
337: PetscErrorCode Device<T>::view_device_(PetscDevice device, PetscViewer viewer) noexcept
338: {
339:   PetscFunctionBegin;
340:   // now this __shouldn't__ reconfigure the device, but there is a petscinfo call to indicate
341:   // it is being reconfigured
342:   PetscCall(devices_[device->deviceId]->configure());
343:   PetscCall(devices_[device->deviceId]->view(viewer));
344:   PetscFunctionReturn(PETSC_SUCCESS);
345: }

347: template <DeviceType T>
348: PetscErrorCode Device<T>::get_attribute_(PetscInt id, PetscDeviceAttribute attr, void *value) noexcept
349: {
350:   PetscFunctionBegin;
351:   PetscCall(devices_[id]->getattribute(attr, value));
352:   PetscFunctionReturn(PETSC_SUCCESS);
353: }

355: // explicitly instantiate the classes
356: #if PetscDefined(HAVE_CUDA)
357: template class Device<DeviceType::CUDA>;
358: #endif
359: #if PetscDefined(HAVE_HIP)
360: template class Device<DeviceType::HIP>;
361: #endif

363: } // namespace cupm

365: } // namespace device

367: } // namespace Petsc