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;

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

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

 71:     PetscCall(CUPMAwareMPI_(&aware));
 72:     // For Open MPI, we could do a compile time check with
 73:     // "defined(PETSC_HAVE_OPENMPI) && defined(MPIX_CUDA_AWARE_SUPPORT) &&
 74:     // MPIX_CUDA_AWARE_SUPPORT" to see if it is CUDA-aware. However, recent versions of IBM
 75:     // Spectrum MPI (e.g., 10.3.1) on Summit meet above conditions, but one has to use jsrun
 76:     // --smpiargs=-gpu to really enable GPU-aware MPI. So we do the check at runtime with a
 77:     // code that works only with GPU-aware MPI.
 78:     if (PetscUnlikely(!aware)) {
 79:       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"));
 80:       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"));
 81:       PetscCall((*PetscErrorPrintf)("If you do care, for IBM Spectrum MPI on OLCF Summit, you may need jsrun --smpiargs=-gpu.\n"));
 82:       PetscCall((*PetscErrorPrintf)("For Open MPI, you need to configure it --with-cuda (https://www.open-mpi.org/faq/?category=buildcuda)\n"));
 83:       PetscCall((*PetscErrorPrintf)("For MVAPICH2-GDR, you need to set MV2_USE_CUDA=1 (http://mvapich.cse.ohio-state.edu/userguide/gdr/)\n"));
 84:       PetscCall((*PetscErrorPrintf)("For Cray-MPICH, you need to set MPICH_GPU_SUPPORT_ENABLED=1 (man mpi to see manual of cray-mpich)\n"));
 85:       PETSCABORT(PETSC_COMM_SELF, PETSC_ERR_LIB);
 86:     }
 87:   }
 88:   PetscFunctionReturn(PETSC_SUCCESS);
 89: }

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

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

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

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

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

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

166: static std::jmp_buf cupmMPIAwareJumpBuffer;
167: static bool         cupmMPIAwareJumpBufferSet;

169: // godspeed to anyone that attempts to call this function
170: void SilenceVariableIsNotNeededAndWillNotBeEmittedWarning_ThisFunctionShouldNeverBeCalled()
171: {
172:   PETSCABORT(MPI_COMM_NULL, (PetscErrorCode)INT_MAX);
173:   if (cupmMPIAwareJumpBufferSet) (void)cupmMPIAwareJumpBuffer;
174: }

176: template <DeviceType T>
177: PetscErrorCode Device<T>::DeviceInternal::CUPMAwareMPI_(bool *awareness) noexcept
178: {
179:   constexpr int hbuf[]            = {1, 0};
180:   int          *dbuf              = nullptr;
181:   const auto    cupmSignalHandler = [](int signal, void *ptr) -> PetscErrorCode {
182:     if ((signal == SIGSEGV) && cupmMPIAwareJumpBufferSet) std::longjmp(cupmMPIAwareJumpBuffer, 1);
183:     return PetscSignalHandlerDefault(signal, ptr);
184:   };

186:   PetscFunctionBegin;
187:   *awareness = false;
188:   PetscCallCUPM(cupmMalloc(reinterpret_cast<void **>(&dbuf), sizeof(hbuf)));
189:   PetscCallCUPM(cupmMemcpy(dbuf, hbuf, sizeof(hbuf), cupmMemcpyHostToDevice));
190:   PetscCallCUPM(cupmDeviceSynchronize());
191:   PetscCall(PetscPushSignalHandler(cupmSignalHandler, nullptr));
192:   cupmMPIAwareJumpBufferSet = true;
193:   if (!setjmp(cupmMPIAwareJumpBuffer) && !MPI_Allreduce(dbuf, dbuf + 1, 1, MPI_INT, MPI_SUM, PETSC_COMM_SELF)) *awareness = true;
194:   cupmMPIAwareJumpBufferSet = false;
195:   PetscCall(PetscPopSignalHandler());
196:   PetscCallCUPM(cupmFree(dbuf));
197:   PetscFunctionReturn(PETSC_SUCCESS);
198: }

200: template <DeviceType T>
201: PetscErrorCode Device<T>::finalize_() noexcept
202: {
203:   PetscFunctionBegin;
204:   if (PetscUnlikely(!initialized_)) PetscFunctionReturn(PETSC_SUCCESS);
205:   for (auto &&device : devices_) device.reset();
206:   defaultDevice_ = PETSC_CUPM_DEVICE_NONE; // disabled by default
207:   initialized_   = false;
208:   PetscFunctionReturn(PETSC_SUCCESS);
209: }

211: template <DeviceType T>
212: PETSC_NODISCARD static PETSC_CONSTEXPR_14 const char *CUPM_VISIBLE_DEVICES() noexcept
213: {
214:   switch (T) {
215:   case DeviceType::CUDA:
216:     return "CUDA_VISIBLE_DEVICES";
217:   case DeviceType::HIP:
218:     return "HIP_VISIBLE_DEVICES";
219:   }
220:   PetscUnreachable();
221:   return "PETSC_ERROR_PLIB";
222: }

224: /*
225:      The default device ID is
226:        MPI     -- rank % number_local_devices
227:        PyTorch -- getenv("LOCAL_RANK")
228: */
229: template <DeviceType T>
230: PetscErrorCode Device<T>::initialize(MPI_Comm comm, PetscInt *defaultDeviceId, PetscBool *defaultView, PetscDeviceInitType *defaultInitType) noexcept
231: {
232:   auto initId   = std::make_pair(*defaultDeviceId, PETSC_FALSE);
233:   auto initView = std::make_pair(*defaultView, PETSC_FALSE);
234:   auto initType = std::make_pair(*defaultInitType, PETSC_FALSE);
235:   int  ndev     = 0;

237:   PetscFunctionBegin;
238:   if (initialized_) PetscFunctionReturn(PETSC_SUCCESS);
239:   initialized_ = true;
240:   PetscCall(PetscRegisterFinalize(finalize_));
241:   PetscCall(base_type::PetscOptionDeviceAll(comm, initType, initId, initView));

243:   if (initType.first == PETSC_DEVICE_INIT_NONE) {
244:     initId.first = PETSC_CUPM_DEVICE_NONE;
245:   } else if (const auto cerr = cupmGetDeviceCount(&ndev)) {
246:     auto PETSC_UNUSED ignored = cupmGetLastError();

248:     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));
249:     // we won't be initializing anything anyways
250:     initType.first = PETSC_DEVICE_INIT_NONE;
251:     // save the error code for later
252:     initId.first = -static_cast<decltype(initId.first)>(cerr);
253:   }

255:   // check again for init type, since the device count may have changed it
256:   if (initType.first == PETSC_DEVICE_INIT_NONE) {
257:     // id < 0 (excluding PETSC_DECIDE) indicates an error has occurred during setup
258:     if ((initId.first > 0) || (initId.first == PETSC_DECIDE)) initId.first = PETSC_CUPM_DEVICE_NONE;
259:     // initType overrides initView
260:     initView.first = PETSC_FALSE;
261:   } else {
262:     PetscCall(PetscDeviceCheckDeviceCount_Internal(ndev));
263:     if (initId.first == PETSC_DECIDE) {
264:       if (ndev) {
265:         /* TORCHELASTIC_RUN_ID is used as a proxy to determine if the current process was launched with torchrun */
266:         char *pytorch_exists = (char *)getenv("TORCHELASTIC_RUN_ID");
267:         char *pytorch_rank   = (char *)getenv("LOCAL_RANK");

269:         if (pytorch_exists && pytorch_rank) {
270:           char *endptr;

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

277:           PetscCallMPI(MPI_Comm_rank(comm, &rank));
278:           initId.first = rank % ndev;
279:         }
280:       } else initId.first = 0;
281:     }
282:     if (initView.first) initType.first = PETSC_DEVICE_INIT_EAGER;
283:   }

285:   static_assert(std::is_same<PetscMPIInt, decltype(defaultDevice_)>::value, "");
286:   // initId.first is PetscInt, _defaultDevice is int
287:   PetscCall(PetscMPIIntCast(initId.first, &defaultDevice_));
288:   // record the results of the initialization
289:   *defaultDeviceId = initId.first;
290:   *defaultView     = initView.first;
291:   *defaultInitType = initType.first;
292:   PetscFunctionReturn(PETSC_SUCCESS);
293: }

295: template <DeviceType T>
296: PetscErrorCode Device<T>::init_device_id_(PetscInt *inid) const noexcept
297: {
298:   const auto id   = *inid == PETSC_DECIDE ? defaultDevice_ : (int)*inid;
299:   const auto cerr = static_cast<cupmError_t>(-defaultDevice_);

301:   PetscFunctionBegin;
302:   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());
303:   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));
304:   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);

306:   if (!devices_[id]) devices_[id] = util::make_unique<DeviceInternal>(id);
307:   PetscAssert(id == devices_[id]->id(), PETSC_COMM_SELF, PETSC_ERR_PLIB, "Entry %d contains device with mismatching id %d", id, devices_[id]->id());
308:   PetscCall(devices_[id]->initialize());
309:   *inid = id;
310:   PetscFunctionReturn(PETSC_SUCCESS);
311: }

313: template <DeviceType T>
314: PetscErrorCode Device<T>::configure_device_(PetscDevice device) noexcept
315: {
316:   PetscFunctionBegin;
317:   PetscCall(devices_[device->deviceId]->configure());
318:   PetscFunctionReturn(PETSC_SUCCESS);
319: }

321: template <DeviceType T>
322: PetscErrorCode Device<T>::view_device_(PetscDevice device, PetscViewer viewer) noexcept
323: {
324:   PetscFunctionBegin;
325:   // now this __shouldn't__ reconfigure the device, but there is a petscinfo call to indicate
326:   // it is being reconfigured
327:   PetscCall(devices_[device->deviceId]->configure());
328:   PetscCall(devices_[device->deviceId]->view(viewer));
329:   PetscFunctionReturn(PETSC_SUCCESS);
330: }

332: template <DeviceType T>
333: PetscErrorCode Device<T>::get_attribute_(PetscInt id, PetscDeviceAttribute attr, void *value) noexcept
334: {
335:   PetscFunctionBegin;
336:   PetscCall(devices_[id]->getattribute(attr, value));
337:   PetscFunctionReturn(PETSC_SUCCESS);
338: }

340: // explicitly instantiate the classes
341: #if PetscDefined(HAVE_CUDA)
342: template class Device<DeviceType::CUDA>;
343: #endif
344: #if PetscDefined(HAVE_HIP)
345: template class Device<DeviceType::HIP>;
346: #endif

348: } // namespace cupm

350: } // namespace device

352: } // namespace Petsc