Actual source code: veccupmimpl.h

  1: #pragma once

  3: #include <petsc/private/vecimpl.h>
  4: #include <../src/vec/vec/impls/dvecimpl.h>

  6: #if PetscDefined(HAVE_NVSHMEM)
  7: PETSC_INTERN PetscErrorCode PetscNvshmemInitializeCheck(void);
  8: PETSC_INTERN PetscErrorCode PetscNvshmemMalloc(size_t, void **);
  9: PETSC_INTERN PetscErrorCode PetscNvshmemCalloc(size_t, void **);
 10: PETSC_INTERN PetscErrorCode PetscNvshmemFree_Private(void *);
 11:   #define PetscNvshmemFree(ptr) ((PetscErrorCode)((ptr) && (PetscNvshmemFree_Private(ptr) || ((ptr) = PETSC_NULLPTR, PETSC_SUCCESS))))
 12: PETSC_INTERN PetscErrorCode PetscNvshmemSum(PetscInt, PetscScalar *, const PetscScalar *);
 13: PETSC_INTERN PetscErrorCode PetscNvshmemMax(PetscInt, PetscReal *, const PetscReal *);
 14: #else
 15:   #define PetscNvshmemFree(ptr) PETSC_SUCCESS
 16: #endif

 18: #if defined(__cplusplus) && PetscDefined(HAVE_DEVICE)
 19: #include <petsc/private/deviceimpl.h>
 20: #include <petsc/private/cupmobject.hpp>
 21: #include <petsc/private/cupmblasinterface.hpp>

 23: #include <petsc/private/cpp/functional.hpp>

 25:   #include <limits> // std::numeric_limits

 27: namespace Petsc
 28: {

 30: namespace vec
 31: {

 33: namespace cupm
 34: {

 36: namespace impl
 37: {

 39: namespace
 40: {

 42: struct no_op {
 43:   template <typename... T>
 44:   constexpr PetscErrorCode operator()(T &&...) const noexcept
 45:   {
 46:     return PETSC_SUCCESS;
 47:   }
 48: };

 50: template <typename T>
 51: struct CooPair {
 52:   using value_type = T;
 53:   using size_type  = PetscCount;

 55:   value_type *&device;
 56:   value_type *&host;
 57:   size_type    size;
 58: };

 60: template <typename U>
 61: static constexpr CooPair<U> make_coo_pair(U *&device, U *&host, PetscCount size) noexcept
 62: {
 63:   return {device, host, size};
 64: }

 66: } // anonymous namespace

 68: // forward declarations
 69: template <device::cupm::DeviceType>
 70: class VecSeq_CUPM;
 71: template <device::cupm::DeviceType>
 72: class VecMPI_CUPM;

 74: // ==========================================================================================
 75: // Vec_CUPMBase
 76: //
 77: // Base class for the VecSeq and VecMPI CUPM implementations. On top of the usual DeviceType
 78: // template parameter it also uses CRTP to be able to use values/calls specific to either
 79: // VecSeq or VecMPI. This is in effect "inside-out" polymorphism.
 80: // ==========================================================================================
 81: template <device::cupm::DeviceType T, typename Derived>
 82: class Vec_CUPMBase : protected device::cupm::impl::CUPMObject<T> {
 83: public:
 84:   PETSC_CUPMOBJECT_HEADER(T);

 86:   // ==========================================================================================
 87:   // Vec_CUPMBase::VectorArray
 88:   //
 89:   // RAII versions of the get/restore array routines. Determines constness of the pointer type,
 90:   // holds the pointer itself provides the implicit conversion operator
 91:   // ==========================================================================================
 92:   template <PetscMemType, PetscMemoryAccessMode>
 93:   class VectorArray;

 95: protected:
 96:   static PetscErrorCode VecView_Debug(Vec v, const char *message = "") noexcept
 97:   {
 98:     const auto   pobj  = PetscObjectCast(v);
 99:     const auto   vimpl = VecIMPLCast(v);
100:     const auto   vcu   = VecCUPMCast(v);
101:     PetscMemType mtype;
102:     MPI_Comm     comm;

104:     PetscFunctionBegin;
105:     PetscAssertPointer(vimpl, 1);
106:     PetscAssertPointer(vcu, 1);
107:     PetscCall(PetscObjectGetComm(pobj, &comm));
108:     PetscCall(PetscPrintf(comm, "---------- %s ----------\n", message));
109:     PetscCall(PetscObjectPrintClassNamePrefixType(pobj, PETSC_VIEWER_STDOUT_(comm)));
110:     PetscCall(PetscPrintf(comm, "Address:             %p\n", v));
111:     PetscCall(PetscPrintf(comm, "Size:                %" PetscInt_FMT "\n", v->map->n));
112:     PetscCall(PetscPrintf(comm, "Offload mask:        %s\n", PetscOffloadMaskToString(v->offloadmask)));
113:     PetscCall(PetscPrintf(comm, "Host ptr:            %p\n", vimpl->array));
114:     PetscCall(PetscPrintf(comm, "Device ptr:          %p\n", vcu->array_d));
115:     PetscCall(PetscPrintf(comm, "Device alloced ptr:  %p\n", vcu->array_allocated_d));
116:     PetscCall(PetscCUPMGetMemType(vcu->array_d, &mtype));
117:     PetscCall(PetscPrintf(comm, "dptr is device mem?  %s\n", PetscBools[static_cast<PetscBool>(PetscMemTypeDevice(mtype))]));
118:     PetscFunctionReturn(PETSC_SUCCESS);
119:   }

121:   // Delete the allocated device array if required and replace it with the given array
122:   static PetscErrorCode ResetAllocatedDevicePtr_(PetscDeviceContext, Vec, PetscScalar * = nullptr) noexcept;
123:   // Check either the host or device impl pointer is allocated and allocate it if
124:   // isn't. CastFunctionType casts the Vec to the required type and returns the pointer
125:   template <typename CastFunctionType>
126:   static PetscErrorCode VecAllocateCheck_(Vec, void *&, CastFunctionType &&) noexcept;
127:   // Check the CUPM part (v->spptr) is allocated, otherwise allocate it
128:   static PetscErrorCode VecCUPMAllocateCheck_(Vec) noexcept;
129:   // Check the Host part (v->data) is allocated, otherwise allocate it
130:   static PetscErrorCode VecIMPLAllocateCheck_(Vec) noexcept;
131:   // Check the Host array is allocated, otherwise allocate it
132:   static PetscErrorCode HostAllocateCheck_(PetscDeviceContext, Vec) noexcept;
133:   // Check the CUPM array is allocated, otherwise allocate it
134:   static PetscErrorCode DeviceAllocateCheck_(PetscDeviceContext, Vec) noexcept;
135:   // Copy HTOD, allocating device if necessary
136:   static PetscErrorCode CopyToDevice_(PetscDeviceContext, Vec, bool = false) noexcept;
137:   // Copy DTOH, allocating host if necessary
138:   static PetscErrorCode CopyToHost_(PetscDeviceContext, Vec, bool = false) noexcept;
139:   static PetscErrorCode DestroyDevice_(Vec) noexcept;
140:   static PetscErrorCode DestroyHost_(Vec) noexcept;

142: public:
143:   struct Vec_CUPM {
144:     PetscScalar *array_d;           // gpu data
145:     PetscScalar *array_allocated_d; // does PETSc own the array ptr?
146:     PetscBool    nvshmem;           // is array allocated in nvshmem? It is used to allocate
147:                                     // Mvctx->lvec in nvshmem

149:     // COO stuff
150:     PetscCount *jmap1_d; // [m+1]: i-th entry of the vector has jmap1[i+1]-jmap1[i] repeats
151:                          // in COO arrays
152:     PetscCount *perm1_d; // [tot1]: permutation array for local entries
153:     PetscCount *imap2_d; // [nnz2]: i-th unique entry in recvbuf is imap2[i]-th entry in
154:                          // the vector
155:     PetscCount *jmap2_d; // [nnz2+1]
156:     PetscCount *perm2_d; // [recvlen]
157:     PetscCount *Cperm_d; // [sendlen]: permutation array to fill sendbuf[]. 'C' for
158:                          // communication

160:     // Buffers for remote values in VecSetValuesCOO()
161:     PetscScalar *sendbuf_d;
162:     PetscScalar *recvbuf_d;
163:   };

165:   // Cast the Vec to its Vec_CUPM struct, i.e. return the result of (Vec_CUPM *)v->spptr
166:   PETSC_NODISCARD static Vec_CUPM *VecCUPMCast(Vec) noexcept;
167:   // Cast the Vec to its host struct, i.e. return the result of (Vec_Seq *)v->data
168:   template <typename U = Derived>
169:   PETSC_NODISCARD static constexpr auto VecIMPLCast(Vec v) noexcept -> decltype(U::VecIMPLCast_(v));
170:   // Get the PetscLogEvents for HTOD and DTOH
171:   PETSC_NODISCARD static constexpr PetscLogEvent VEC_CUPMCopyToGPU() noexcept;
172:   PETSC_NODISCARD static constexpr PetscLogEvent VEC_CUPMCopyFromGPU() noexcept;
173:   // Get the VecTypes
174:   PETSC_NODISCARD static constexpr VecType VECSEQCUPM() noexcept;
175:   PETSC_NODISCARD static constexpr VecType VECMPICUPM() noexcept;
176:   PETSC_NODISCARD static constexpr VecType VECCUPM() noexcept;

178:   // Get the device VecType of the calling vector
179:   template <typename U = Derived>
180:   PETSC_NODISCARD static constexpr VecType VECIMPLCUPM() noexcept;
181:   // Get the host VecType of the calling vector
182:   template <typename U = Derived>
183:   PETSC_NODISCARD static constexpr VecType VECIMPL() noexcept;

185:   // Call the host destroy function, i.e. VecDestroy_Seq()
186:   static PetscErrorCode VecDestroy_IMPL(Vec) noexcept;
187:   // Call the host reset function, i.e. VecResetArray_Seq()
188:   static PetscErrorCode VecResetArray_IMPL(Vec) noexcept;
189:   // ... you get the idea
190:   static PetscErrorCode VecPlaceArray_IMPL(Vec, const PetscScalar *) noexcept;
191:   // Call the host creation function, i.e. VecCreate_Seq(), and also initialize the CUPM part
192:   // along with it if needed
193:   static PetscErrorCode VecCreate_IMPL_Private(Vec, PetscBool *, PetscInt = 0, PetscScalar * = nullptr) noexcept;

195:   // Shorthand for creating VectorArray's. Need functions to create them, otherwise using them
196:   // as an unnamed temporary leads to most vexing parse
197:   PETSC_NODISCARD static auto DeviceArrayRead(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_DEVICE, PETSC_MEMORY_ACCESS_READ>{dctx, v});
198:   PETSC_NODISCARD static auto DeviceArrayWrite(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_DEVICE, PETSC_MEMORY_ACCESS_WRITE>{dctx, v});
199:   PETSC_NODISCARD static auto DeviceArrayReadWrite(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_DEVICE, PETSC_MEMORY_ACCESS_READ_WRITE>{dctx, v});
200:   PETSC_NODISCARD static auto HostArrayRead(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ>{dctx, v});
201:   PETSC_NODISCARD static auto HostArrayWrite(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_WRITE>{dctx, v});
202:   PETSC_NODISCARD static auto HostArrayReadWrite(PetscDeviceContext dctx, Vec v) noexcept PETSC_DECLTYPE_AUTO_RETURNS(VectorArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ_WRITE>{dctx, v});

204:   // ops-table functions
205:   static PetscErrorCode Create(Vec) noexcept;
206:   static PetscErrorCode Destroy(Vec) noexcept;
207:   template <PetscMemType, PetscMemoryAccessMode, bool = false>
208:   static PetscErrorCode GetArray(Vec, PetscScalar **, PetscDeviceContext) noexcept;
209:   template <PetscMemType, PetscMemoryAccessMode, bool = false>
210:   static PetscErrorCode GetArray(Vec, PetscScalar **) noexcept;
211:   template <PetscMemType, PetscMemoryAccessMode>
212:   static PetscErrorCode RestoreArray(Vec, PetscScalar **, PetscDeviceContext) noexcept;
213:   template <PetscMemType, PetscMemoryAccessMode>
214:   static PetscErrorCode RestoreArray(Vec, PetscScalar **) noexcept;
215:   template <PetscMemoryAccessMode>
216:   static PetscErrorCode GetArrayAndMemtype(Vec, PetscScalar **, PetscMemType *, PetscDeviceContext) noexcept;
217:   template <PetscMemoryAccessMode>
218:   static PetscErrorCode GetArrayAndMemtype(Vec, PetscScalar **, PetscMemType *) noexcept;
219:   template <PetscMemoryAccessMode>
220:   static PetscErrorCode RestoreArrayAndMemtype(Vec, PetscScalar **, PetscDeviceContext) noexcept;
221:   template <PetscMemoryAccessMode>
222:   static PetscErrorCode RestoreArrayAndMemtype(Vec, PetscScalar **) noexcept;
223:   template <PetscMemType>
224:   static PetscErrorCode ReplaceArray(Vec, const PetscScalar *) noexcept;
225:   template <PetscMemType>
226:   static PetscErrorCode ResetArray(Vec) noexcept;
227:   template <PetscMemType>
228:   static PetscErrorCode PlaceArray(Vec, const PetscScalar *) noexcept;

230:   // common ops shared between Seq and MPI
231:   static PetscErrorCode Create_CUPM(Vec) noexcept;
232:   static PetscErrorCode Create_CUPMBase(MPI_Comm, PetscInt, PetscInt, PetscInt, Vec *, PetscBool, PetscLayout /*reference*/ = nullptr) noexcept;
233:   static PetscErrorCode Initialize_CUPMBase(Vec, PetscBool, PetscScalar *, PetscScalar *, PetscDeviceContext) noexcept;
234:   template <typename SetupFunctionT = no_op>
235:   static PetscErrorCode Duplicate_CUPMBase(Vec, Vec *, PetscDeviceContext, SetupFunctionT && = SetupFunctionT{}) noexcept;
236:   static PetscErrorCode BindToCPU_CUPMBase(Vec, PetscBool, PetscDeviceContext) noexcept;
237:   static PetscErrorCode GetArrays_CUPMBase(Vec, const PetscScalar **, const PetscScalar **, PetscOffloadMask *, PetscDeviceContext) noexcept;
238:   static PetscErrorCode ResetPreallocationCOO_CUPMBase(Vec, PetscDeviceContext) noexcept;
239:   template <std::size_t NCount = 0, std::size_t NScal = 0>
240:   static PetscErrorCode SetPreallocationCOO_CUPMBase(Vec, PetscCount, const PetscInt[], PetscDeviceContext, const std::array<CooPair<PetscCount>, NCount> & = {}, const std::array<CooPair<PetscScalar>, NScal> & = {}) noexcept;

242:   static PetscErrorCode Convert_IMPL_IMPLCUPM(Vec) noexcept;
243: };

245: // ==========================================================================================
246: // Vec_CUPMBase::VectorArray
247: //
248: // RAII versions of the get/restore array routines. Determines constness of the pointer type,
249: // holds the pointer itself and provides the implicit conversion operator.
250: //
251: // On construction this calls the moral equivalent of Vec[CUPM]GetArray[Read|Write]()
252: // (depending on PetscMemoryAccessMode) and on destruction automatically restores the array
253: // for you
254: // ==========================================================================================
255: template <device::cupm::DeviceType T, typename D>
256: template <PetscMemType MT, PetscMemoryAccessMode MA>
257: class Vec_CUPMBase<T, D>::VectorArray : public device::cupm::impl::RestoreableArray<T, MT, MA> {
258:   using base_type = device::cupm::impl::RestoreableArray<T, MT, MA>;

260: public:
261:   VectorArray(PetscDeviceContext, Vec) noexcept;
262:   ~VectorArray() noexcept;

264: private:
265:   Vec v_ = nullptr;
266: };

268: // ==========================================================================================
269: // Vec_CUPMBase::VectorArray - Public API
270: // ==========================================================================================

272: template <device::cupm::DeviceType T, typename D>
273: template <PetscMemType MT, PetscMemoryAccessMode MA>
274: inline Vec_CUPMBase<T, D>::VectorArray<MT, MA>::VectorArray(PetscDeviceContext dctx, Vec v) noexcept : base_type{dctx}, v_{v}
275: {
276:   PetscFunctionBegin;
277:   PetscCallAbort(PETSC_COMM_SELF, Vec_CUPMBase<T, D>::template GetArray<MT, MA, true>(v, &this->ptr_, dctx));
278:   PetscFunctionReturnVoid();
279: }

281: template <device::cupm::DeviceType T, typename D>
282: template <PetscMemType MT, PetscMemoryAccessMode MA>
283: inline Vec_CUPMBase<T, D>::VectorArray<MT, MA>::~VectorArray() noexcept
284: {
285:   PetscFunctionBegin;
286:   PetscCallAbort(PETSC_COMM_SELF, Vec_CUPMBase<T, D>::template RestoreArray<MT, MA>(v_, &this->ptr_, this->dctx_));
287:   PetscFunctionReturnVoid();
288: }

290: // ==========================================================================================
291: // Vec_CUPMBase - Protected API
292: // ==========================================================================================

294: template <device::cupm::DeviceType T, typename D>
295: inline PetscErrorCode Vec_CUPMBase<T, D>::ResetAllocatedDevicePtr_(PetscDeviceContext dctx, Vec v, PetscScalar *new_value) noexcept
296: {
297:   auto &device_array = VecCUPMCast(v)->array_allocated_d;

299:   PetscFunctionBegin;
300:   if (device_array) {
301:     if (PetscDefined(HAVE_NVSHMEM) && VecCUPMCast(v)->nvshmem) {
302:       PetscCall(PetscNvshmemFree(device_array));
303:     } else {
304:       cupmStream_t stream;

306:       PetscCall(GetHandlesFrom_(dctx, &stream));
307:       PetscCallCUPM(cupmFreeAsync(device_array, stream));
308:     }
309:   }
310:   device_array = new_value;
311:   PetscFunctionReturn(PETSC_SUCCESS);
312: }

314: namespace
315: {

317: inline PetscErrorCode VecCUPMCheckMinimumPinnedMemory_Internal(Vec v, PetscBool *set = nullptr) noexcept
318: {
319:   auto      mem = static_cast<PetscInt>(v->minimum_bytes_pinned_memory);
320:   PetscBool flg;

322:   PetscFunctionBegin;
323:   PetscObjectOptionsBegin(PetscObjectCast(v));
324:   PetscCall(PetscOptionsRangeInt("-vec_pinned_memory_min", "Minimum size (in bytes) for an allocation to use pinned memory on host", "VecSetPinnedMemoryMin", mem, &mem, &flg, 0, std::numeric_limits<decltype(mem)>::max()));
325:   if (flg) v->minimum_bytes_pinned_memory = mem;
326:   PetscOptionsEnd();
327:   if (set) *set = flg;
328:   PetscFunctionReturn(PETSC_SUCCESS);
329: }

331: } // anonymous namespace

333: template <device::cupm::DeviceType T, typename D>
334: template <typename CastFunctionType>
335: inline PetscErrorCode Vec_CUPMBase<T, D>::VecAllocateCheck_(Vec v, void *&dest, CastFunctionType &&cast) noexcept
336: {
337:   PetscFunctionBegin;
338:   if (PetscLikely(dest)) PetscFunctionReturn(PETSC_SUCCESS);
339:   // do the check here so we don't have to do it in every function
340:   PetscCall(checkCupmBlasIntCast(v->map->n));
341:   {
342:     auto impl = cast(v);

344:     PetscCall(PetscNew(&impl));
345:     dest = impl;
346:   }
347:   PetscFunctionReturn(PETSC_SUCCESS);
348: }

350: template <device::cupm::DeviceType T, typename D>
351: inline PetscErrorCode Vec_CUPMBase<T, D>::VecIMPLAllocateCheck_(Vec v) noexcept
352: {
353:   PetscFunctionBegin;
354:   PetscCall(VecAllocateCheck_(v, v->data, VecIMPLCast<D>));
355:   PetscFunctionReturn(PETSC_SUCCESS);
356: }

358: // allocate the Vec_CUPM struct. this is normally done through DeviceAllocateCheck_(), but in
359: // certain circumstances (such as when the user places the device array) we do not want to do
360: // the full DeviceAllocateCheck_() as it also allocates the array
361: template <device::cupm::DeviceType T, typename D>
362: inline PetscErrorCode Vec_CUPMBase<T, D>::VecCUPMAllocateCheck_(Vec v) noexcept
363: {
364:   PetscFunctionBegin;
365:   PetscCall(VecAllocateCheck_(v, v->spptr, VecCUPMCast));
366:   PetscFunctionReturn(PETSC_SUCCESS);
367: }

369: template <device::cupm::DeviceType T, typename D>
370: inline PetscErrorCode Vec_CUPMBase<T, D>::HostAllocateCheck_(PetscDeviceContext, Vec v) noexcept
371: {
372:   PetscFunctionBegin;
373:   PetscCall(VecIMPLAllocateCheck_(v));
374:   if (auto &alloc = VecIMPLCast(v)->array_allocated) PetscFunctionReturn(PETSC_SUCCESS);
375:   else {
376:     PetscCall(VecCUPMCheckMinimumPinnedMemory_Internal(v));
377:     {
378:       const auto n     = v->map->n;
379:       const auto useit = UseCUPMHostAlloc((n * sizeof(*alloc)) > v->minimum_bytes_pinned_memory);

381:       v->pinned_memory = static_cast<decltype(v->pinned_memory)>(useit.value());
382:       PetscCall(PetscMalloc1(n, &alloc));
383:     }
384:     if (!VecIMPLCast(v)->array) VecIMPLCast(v)->array = alloc;
385:     if (v->offloadmask == PETSC_OFFLOAD_UNALLOCATED) v->offloadmask = PETSC_OFFLOAD_CPU;
386:   }
387:   PetscFunctionReturn(PETSC_SUCCESS);
388: }

390: template <device::cupm::DeviceType T, typename D>
391: inline PetscErrorCode Vec_CUPMBase<T, D>::DeviceAllocateCheck_(PetscDeviceContext dctx, Vec v) noexcept
392: {
393:   PetscFunctionBegin;
394:   PetscCall(VecCUPMAllocateCheck_(v));
395:   if (auto &alloc = VecCUPMCast(v)->array_d) PetscFunctionReturn(PETSC_SUCCESS);
396:   else {
397:     const auto   n                 = v->map->n;
398:     auto        &array_allocated_d = VecCUPMCast(v)->array_allocated_d;
399:     cupmStream_t stream;

401:     PetscCall(GetHandlesFrom_(dctx, &stream));
402:     PetscCall(PetscCUPMMallocAsync(&array_allocated_d, n, stream));
403:     alloc = array_allocated_d;
404:     if (v->offloadmask == PETSC_OFFLOAD_UNALLOCATED) {
405:       const auto vimp = VecIMPLCast(v);
406:       v->offloadmask  = (vimp && vimp->array) ? PETSC_OFFLOAD_CPU : PETSC_OFFLOAD_GPU;
407:     }
408:   }
409:   PetscFunctionReturn(PETSC_SUCCESS);
410: }

412: template <device::cupm::DeviceType T, typename D>
413: inline PetscErrorCode Vec_CUPMBase<T, D>::CopyToDevice_(PetscDeviceContext dctx, Vec v, bool forceasync) noexcept
414: {
415:   PetscFunctionBegin;
416:   PetscCall(DeviceAllocateCheck_(dctx, v));
417:   if (v->offloadmask == PETSC_OFFLOAD_CPU) {
418:     cupmStream_t stream;

420:     v->offloadmask = PETSC_OFFLOAD_BOTH;
421:     PetscCall(GetHandlesFrom_(dctx, &stream));
422:     PetscCall(PetscLogEventBegin(VEC_CUPMCopyToGPU(), v, 0, 0, 0));
423:     PetscCall(PetscCUPMMemcpyAsync(VecCUPMCast(v)->array_d, VecIMPLCast(v)->array, v->map->n, cupmMemcpyHostToDevice, stream, forceasync));
424:     PetscCall(PetscLogEventEnd(VEC_CUPMCopyToGPU(), v, 0, 0, 0));
425:   }
426:   PetscFunctionReturn(PETSC_SUCCESS);
427: }

429: template <device::cupm::DeviceType T, typename D>
430: inline PetscErrorCode Vec_CUPMBase<T, D>::CopyToHost_(PetscDeviceContext dctx, Vec v, bool forceasync) noexcept
431: {
432:   PetscFunctionBegin;
433:   PetscCall(HostAllocateCheck_(dctx, v));
434:   if (v->offloadmask == PETSC_OFFLOAD_GPU) {
435:     cupmStream_t stream;

437:     v->offloadmask = PETSC_OFFLOAD_BOTH;
438:     PetscCall(GetHandlesFrom_(dctx, &stream));
439:     PetscCall(PetscLogEventBegin(VEC_CUPMCopyFromGPU(), v, 0, 0, 0));
440:     PetscCall(PetscCUPMMemcpyAsync(VecIMPLCast(v)->array, VecCUPMCast(v)->array_d, v->map->n, cupmMemcpyDeviceToHost, stream, forceasync));
441:     PetscCall(PetscLogEventEnd(VEC_CUPMCopyFromGPU(), v, 0, 0, 0));
442:   }
443:   PetscFunctionReturn(PETSC_SUCCESS);
444: }

446: template <device::cupm::DeviceType T, typename D>
447: inline PetscErrorCode Vec_CUPMBase<T, D>::DestroyDevice_(Vec v) noexcept
448: {
449:   PetscFunctionBegin;
450:   if (const auto vcu = VecCUPMCast(v)) {
451:     PetscDeviceContext dctx;

453:     PetscCall(GetHandles_(&dctx));
454:     PetscCall(ResetAllocatedDevicePtr_(dctx, v));
455:     PetscCall(ResetPreallocationCOO_CUPMBase(v, dctx));
456:     PetscCall(PetscFree(v->spptr));
457:   }
458:   PetscFunctionReturn(PETSC_SUCCESS);
459: }

461: template <device::cupm::DeviceType T, typename D>
462: inline PetscErrorCode Vec_CUPMBase<T, D>::DestroyHost_(Vec v) noexcept
463: {
464:   PetscFunctionBegin;
465:   PetscCall(PetscObjectSAWsViewOff(PetscObjectCast(v)));
466:   if (const auto vimpl = VecIMPLCast(v)) {
467:     if (auto &array_allocated = vimpl->array_allocated) {
468:       const auto useit = UseCUPMHostAlloc(v->pinned_memory);

470:       // do this ourselves since we may want to use the cupm functions
471:       PetscCall(PetscFree(array_allocated));
472:     }
473:   }
474:   v->pinned_memory = PETSC_FALSE;
475:   PetscCall(VecDestroy_IMPL(v));
476:   PetscFunctionReturn(PETSC_SUCCESS);
477: }

479: // ==========================================================================================
480: // Vec_CUPMBase - Public API
481: // ==========================================================================================

483: template <device::cupm::DeviceType T, typename D>
484: inline typename Vec_CUPMBase<T, D>::Vec_CUPM *Vec_CUPMBase<T, D>::VecCUPMCast(Vec v) noexcept
485: {
486:   return static_cast<Vec_CUPM *>(v->spptr);
487: }

489: // This is a trick to get around the fact that in CRTP the derived class is not yet fully
490: // defined because Base must necessarily be instantiated before Derived is
491: // complete. By using a dummy template parameter we make the type "dependent" and so will
492: // only be determined when the derived class is instantiated (and therefore fully defined)
493: template <device::cupm::DeviceType T, typename D>
494: template <typename U>
495: inline constexpr auto Vec_CUPMBase<T, D>::VecIMPLCast(Vec v) noexcept -> decltype(U::VecIMPLCast_(v))
496: {
497:   return U::VecIMPLCast_(v);
498: }

500: template <device::cupm::DeviceType T, typename D>
501: inline PetscErrorCode Vec_CUPMBase<T, D>::VecDestroy_IMPL(Vec v) noexcept
502: {
503:   return D::VecDestroy_IMPL_(v);
504: }

506: template <device::cupm::DeviceType T, typename D>
507: inline PetscErrorCode Vec_CUPMBase<T, D>::VecResetArray_IMPL(Vec v) noexcept
508: {
509:   return D::VecResetArray_IMPL_(v);
510: }

512: template <device::cupm::DeviceType T, typename D>
513: inline PetscErrorCode Vec_CUPMBase<T, D>::VecPlaceArray_IMPL(Vec v, const PetscScalar *a) noexcept
514: {
515:   return D::VecPlaceArray_IMPL_(v, a);
516: }

518: template <device::cupm::DeviceType T, typename D>
519: inline PetscErrorCode Vec_CUPMBase<T, D>::VecCreate_IMPL_Private(Vec v, PetscBool *alloc_missing, PetscInt nghost, PetscScalar *host_array) noexcept
520: {
521:   return D::VecCreate_IMPL_Private_(v, alloc_missing, nghost, host_array);
522: }

524: template <device::cupm::DeviceType T, typename D>
525: inline constexpr PetscLogEvent Vec_CUPMBase<T, D>::VEC_CUPMCopyToGPU() noexcept
526: {
527:   return T == device::cupm::DeviceType::CUDA ? VEC_CUDACopyToGPU : VEC_HIPCopyToGPU;
528: }

530: template <device::cupm::DeviceType T, typename D>
531: inline constexpr PetscLogEvent Vec_CUPMBase<T, D>::VEC_CUPMCopyFromGPU() noexcept
532: {
533:   return T == device::cupm::DeviceType::CUDA ? VEC_CUDACopyFromGPU : VEC_HIPCopyFromGPU;
534: }

536: template <device::cupm::DeviceType T, typename D>
537: inline constexpr VecType Vec_CUPMBase<T, D>::VECSEQCUPM() noexcept
538: {
539:   return T == device::cupm::DeviceType::CUDA ? VECSEQCUDA : VECSEQHIP;
540: }

542: template <device::cupm::DeviceType T, typename D>
543: inline constexpr VecType Vec_CUPMBase<T, D>::VECMPICUPM() noexcept
544: {
545:   return T == device::cupm::DeviceType::CUDA ? VECMPICUDA : VECMPIHIP;
546: }

548: template <device::cupm::DeviceType T, typename D>
549: inline constexpr VecType Vec_CUPMBase<T, D>::VECCUPM() noexcept
550: {
551:   return T == device::cupm::DeviceType::CUDA ? VECCUDA : VECHIP;
552: }

554: template <device::cupm::DeviceType T, typename D>
555: template <typename U>
556: inline constexpr VecType Vec_CUPMBase<T, D>::VECIMPLCUPM() noexcept
557: {
558:   return U::VECIMPLCUPM_();
559: }

561: template <device::cupm::DeviceType T, typename D>
562: template <typename U>
563: inline constexpr VecType Vec_CUPMBase<T, D>::VECIMPL() noexcept
564: {
565:   return U::VECIMPL_();
566: }

568: // private version that takes a PetscDeviceContext, called by the public variant
569: template <device::cupm::DeviceType T, typename D>
570: template <PetscMemType mtype, PetscMemoryAccessMode access, bool force>
571: inline PetscErrorCode Vec_CUPMBase<T, D>::GetArray(Vec v, PetscScalar **a, PetscDeviceContext dctx) noexcept
572: {
573:   constexpr auto hostmem     = PetscMemTypeHost(mtype);
574:   const auto     oldmask     = v->offloadmask;
575:   auto          &mask        = v->offloadmask;
576:   auto           should_sync = false;

578:   PetscFunctionBegin;
579:   static_assert((mtype == PETSC_MEMTYPE_HOST) || (mtype == PETSC_MEMTYPE_DEVICE), "");
580:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
581:   if (PetscMemoryAccessRead(access)) {
582:     // READ or READ_WRITE
583:     if (((oldmask == PETSC_OFFLOAD_GPU) && hostmem) || ((oldmask == PETSC_OFFLOAD_CPU) && !hostmem)) {
584:       // if we move the data we should set the flag to synchronize later on
585:       should_sync = true;
586:     }
587:     PetscCall((hostmem ? CopyToHost_ : CopyToDevice_)(dctx, v, force));
588:   } else {
589:     // WRITE only
590:     PetscCall((hostmem ? HostAllocateCheck_ : DeviceAllocateCheck_)(dctx, v));
591:   }
592:   *a = hostmem ? VecIMPLCast(v)->array : VecCUPMCast(v)->array_d;
593:   // if unallocated previously we should zero things out if we intend to read
594:   if (PetscMemoryAccessRead(access) && (oldmask == PETSC_OFFLOAD_UNALLOCATED)) {
595:     const auto n = v->map->n;

597:     if (hostmem) {
598:       PetscCall(PetscArrayzero(*a, n));
599:     } else {
600:       cupmStream_t stream;

602:       PetscCall(GetHandlesFrom_(dctx, &stream));
603:       PetscCall(PetscCUPMMemsetAsync(*a, 0, n, stream, force));
604:       should_sync = true;
605:     }
606:   }
607:   // update the offloadmask if we intend to write, since we assume immediately modified
608:   if (PetscMemoryAccessWrite(access)) {
609:     PetscCall(VecSetErrorIfLocked(v, 1));
610:     // REVIEW ME: this should probably also call PetscObjectStateIncrease() since we assume it
611:     // is immediately modified
612:     mask = hostmem ? PETSC_OFFLOAD_CPU : PETSC_OFFLOAD_GPU;
613:   }
614:   // if we are a globally blocking stream and we have MOVED data then we should synchronize,
615:   // since even doing async calls on the NULL stream is not synchronous
616:   if (!force && should_sync) PetscCall(PetscDeviceContextSynchronize(dctx));
617:   PetscFunctionReturn(PETSC_SUCCESS);
618: }

620: // v->ops->getarray[read|write] or VecCUPMGetArray[Read|Write]()
621: template <device::cupm::DeviceType T, typename D>
622: template <PetscMemType mtype, PetscMemoryAccessMode access, bool force>
623: inline PetscErrorCode Vec_CUPMBase<T, D>::GetArray(Vec v, PetscScalar **a) noexcept
624: {
625:   PetscDeviceContext dctx;

627:   PetscFunctionBegin;
628:   PetscCall(GetHandles_(&dctx));
629:   PetscCall(D::template GetArray<mtype, access, force>(v, a, dctx));
630:   PetscFunctionReturn(PETSC_SUCCESS);
631: }

633: // private version that takes a PetscDeviceContext, called by the public variant
634: template <device::cupm::DeviceType T, typename D>
635: template <PetscMemType mtype, PetscMemoryAccessMode access>
636: inline PetscErrorCode Vec_CUPMBase<T, D>::RestoreArray(Vec v, PetscScalar **a, PetscDeviceContext) noexcept
637: {
638:   PetscFunctionBegin;
639:   static_assert((mtype == PETSC_MEMTYPE_HOST) || (mtype == PETSC_MEMTYPE_DEVICE), "");
640:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
641:   if (PetscMemoryAccessWrite(access)) {
642:     // WRITE or READ_WRITE
643:     PetscCall(PetscObjectStateIncrease(PetscObjectCast(v)));
644:     v->offloadmask = PetscMemTypeHost(mtype) ? PETSC_OFFLOAD_CPU : PETSC_OFFLOAD_GPU;
645:   }
646:   if (a) {
647:     PetscCall(CheckPointerMatchesMemType_(*a, mtype));
648:     *a = nullptr;
649:   }
650:   PetscFunctionReturn(PETSC_SUCCESS);
651: }

653: // v->ops->restorearray[read|write] or VecCUPMRestoreArray[Read|Write]()
654: template <device::cupm::DeviceType T, typename D>
655: template <PetscMemType mtype, PetscMemoryAccessMode access>
656: inline PetscErrorCode Vec_CUPMBase<T, D>::RestoreArray(Vec v, PetscScalar **a) noexcept
657: {
658:   PetscDeviceContext dctx;

660:   PetscFunctionBegin;
661:   PetscCall(GetHandles_(&dctx));
662:   PetscCall(D::template RestoreArray<mtype, access>(v, a, dctx));
663:   PetscFunctionReturn(PETSC_SUCCESS);
664: }

666: template <device::cupm::DeviceType T, typename D>
667: template <PetscMemoryAccessMode access>
668: inline PetscErrorCode Vec_CUPMBase<T, D>::GetArrayAndMemtype(Vec v, PetscScalar **a, PetscMemType *mtype, PetscDeviceContext dctx) noexcept
669: {
670:   PetscFunctionBegin;
671:   if (a) PetscCall(D::template GetArray<PETSC_MEMTYPE_DEVICE, access>(v, a, dctx));
672:   if (mtype) *mtype = (PetscDefined(HAVE_NVSHMEM) && VecCUPMCast(v)->nvshmem) ? PETSC_MEMTYPE_NVSHMEM : PETSC_MEMTYPE_CUPM();
673:   PetscFunctionReturn(PETSC_SUCCESS);
674: }

676: // v->ops->getarrayandmemtype
677: template <device::cupm::DeviceType T, typename D>
678: template <PetscMemoryAccessMode access>
679: inline PetscErrorCode Vec_CUPMBase<T, D>::GetArrayAndMemtype(Vec v, PetscScalar **a, PetscMemType *mtype) noexcept
680: {
681:   PetscDeviceContext dctx;

683:   PetscFunctionBegin;
684:   PetscCall(GetHandles_(&dctx));
685:   PetscCall(D::template GetArrayAndMemtype<access>(v, a, mtype, dctx));
686:   PetscFunctionReturn(PETSC_SUCCESS);
687: }

689: template <device::cupm::DeviceType T, typename D>
690: template <PetscMemoryAccessMode access>
691: inline PetscErrorCode Vec_CUPMBase<T, D>::RestoreArrayAndMemtype(Vec v, PetscScalar **a, PetscDeviceContext dctx) noexcept
692: {
693:   PetscFunctionBegin;
694:   PetscCall(D::template RestoreArray<PETSC_MEMTYPE_DEVICE, access>(v, a, dctx));
695:   PetscFunctionReturn(PETSC_SUCCESS);
696: }

698: // v->ops->restorearrayandmemtype
699: template <device::cupm::DeviceType T, typename D>
700: template <PetscMemoryAccessMode access>
701: inline PetscErrorCode Vec_CUPMBase<T, D>::RestoreArrayAndMemtype(Vec v, PetscScalar **a) noexcept
702: {
703:   PetscDeviceContext dctx;

705:   PetscFunctionBegin;
706:   PetscCall(GetHandles_(&dctx));
707:   PetscCall(D::template RestoreArrayAndMemtype<access>(v, a, dctx));
708:   PetscFunctionReturn(PETSC_SUCCESS);
709: }

711: // v->ops->placearray or VecCUPMPlaceArray()
712: template <device::cupm::DeviceType T, typename D>
713: template <PetscMemType mtype>
714: inline PetscErrorCode Vec_CUPMBase<T, D>::PlaceArray(Vec v, const PetscScalar *a) noexcept
715: {
716:   PetscDeviceContext dctx;

718:   PetscFunctionBegin;
719:   static_assert((mtype == PETSC_MEMTYPE_HOST) || (mtype == PETSC_MEMTYPE_DEVICE), "");
720:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
721:   PetscCall(CheckPointerMatchesMemType_(a, mtype));
722:   PetscCall(GetHandles_(&dctx));
723:   if (PetscMemTypeHost(mtype)) {
724:     PetscCall(CopyToHost_(dctx, v));
725:     PetscCall(VecPlaceArray_IMPL(v, a));
726:     v->offloadmask = PETSC_OFFLOAD_CPU;
727:   } else {
728:     PetscCall(VecIMPLAllocateCheck_(v));
729:     {
730:       auto &backup_array = VecIMPLCast(v)->unplacedarray;

732:       PetscCheck(!backup_array, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "VecPlaceArray() was already called on this vector, without a call to VecResetArray()");
733:       PetscCall(CopyToDevice_(dctx, v));
734:       PetscCall(PetscObjectStateIncrease(PetscObjectCast(v)));
735:       backup_array = util::exchange(VecCUPMCast(v)->array_d, const_cast<PetscScalar *>(a));
736:       // only update the offload mask if we actually assign a pointer
737:       if (a) v->offloadmask = PETSC_OFFLOAD_GPU;
738:     }
739:   }
740:   PetscFunctionReturn(PETSC_SUCCESS);
741: }

743: // v->ops->replacearray or VecCUPMReplaceArray()
744: template <device::cupm::DeviceType T, typename D>
745: template <PetscMemType mtype>
746: inline PetscErrorCode Vec_CUPMBase<T, D>::ReplaceArray(Vec v, const PetscScalar *a) noexcept
747: {
748:   const auto         aptr = const_cast<PetscScalar *>(a);
749:   PetscDeviceContext dctx;

751:   PetscFunctionBegin;
752:   static_assert((mtype == PETSC_MEMTYPE_HOST) || (mtype == PETSC_MEMTYPE_DEVICE), "");
753:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
754:   PetscCall(CheckPointerMatchesMemType_(a, mtype));
755:   PetscCall(GetHandles_(&dctx));
756:   if (PetscMemTypeHost(mtype)) {
757:     PetscCall(VecIMPLAllocateCheck_(v));
758:     {
759:       const auto vimpl      = VecIMPLCast(v);
760:       auto      &host_array = vimpl->array_allocated;

762:       // make sure the users array has the latest values.
763:       // REVIEW ME: why? we're about to free it
764:       if (host_array != vimpl->array) PetscCall(CopyToHost_(dctx, v));
765:       if (host_array) {
766:         const auto useit = UseCUPMHostAlloc(v->pinned_memory);

768:         PetscCall(PetscFree(host_array));
769:       }
770:       host_array       = aptr;
771:       vimpl->array     = host_array;
772:       v->pinned_memory = PETSC_FALSE; // REVIEW ME: we can determine this
773:       v->offloadmask   = PETSC_OFFLOAD_CPU;
774:     }
775:   } else {
776:     PetscCall(VecCUPMAllocateCheck_(v));
777:     {
778:       const auto vcu = VecCUPMCast(v);

780:       PetscCall(ResetAllocatedDevicePtr_(dctx, v, aptr));
781:       // don't update the offloadmask if placed pointer is NULL
782:       vcu->array_d = vcu->array_allocated_d /* = aptr */;
783:       if (aptr) v->offloadmask = PETSC_OFFLOAD_GPU;
784:     }
785:   }
786:   PetscCall(PetscObjectStateIncrease(PetscObjectCast(v)));
787:   PetscFunctionReturn(PETSC_SUCCESS);
788: }

790: // v->ops->resetarray or VecCUPMResetArray()
791: template <device::cupm::DeviceType T, typename D>
792: template <PetscMemType mtype>
793: inline PetscErrorCode Vec_CUPMBase<T, D>::ResetArray(Vec v) noexcept
794: {
795:   PetscDeviceContext dctx;

797:   PetscFunctionBegin;
798:   static_assert((mtype == PETSC_MEMTYPE_HOST) || (mtype == PETSC_MEMTYPE_DEVICE), "");
799:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
800:   PetscCall(GetHandles_(&dctx));
801:   // REVIEW ME:
802:   // this is wildly inefficient but must be done if we assume that the placed array must have
803:   // correct values
804:   if (PetscMemTypeHost(mtype)) {
805:     PetscCall(CopyToHost_(dctx, v));
806:     PetscCall(VecResetArray_IMPL(v));
807:     v->offloadmask = PETSC_OFFLOAD_CPU;
808:   } else {
809:     PetscCall(VecIMPLAllocateCheck_(v));
810:     PetscCall(VecCUPMAllocateCheck_(v));
811:     {
812:       const auto vcu        = VecCUPMCast(v);
813:       const auto vimpl      = VecIMPLCast(v);
814:       auto      &host_array = vimpl->unplacedarray;

816:       PetscCall(CheckPointerMatchesMemType_(host_array, PETSC_MEMTYPE_DEVICE));
817:       if (v->offloadmask == PETSC_OFFLOAD_CPU) {
818:         PetscCall(CopyToDevice_(dctx, v));
819:         PetscCall(PetscDeviceContextSynchronize(dctx)); // Above H2D might be async, so we must sync dctx, otherwise if later user writes v's host array, it could ruin the H2D
820:       }
821:       PetscCall(PetscObjectStateIncrease(PetscObjectCast(v)));
822:       // Need to reset the offloadmask. If we had a stashed pointer we are on the GPU,
823:       // otherwise check if the host has a valid pointer. If neither, then we are not
824:       // allocated.
825:       vcu->array_d = host_array;
826:       if (host_array) {
827:         host_array     = nullptr;
828:         v->offloadmask = PETSC_OFFLOAD_GPU;
829:       } else if (vimpl->array) {
830:         v->offloadmask = PETSC_OFFLOAD_CPU;
831:       } else {
832:         v->offloadmask = PETSC_OFFLOAD_UNALLOCATED;
833:       }
834:     }
835:   }
836:   PetscFunctionReturn(PETSC_SUCCESS);
837: }

839: // v->ops->create
840: template <device::cupm::DeviceType T, typename D>
841: inline PetscErrorCode Vec_CUPMBase<T, D>::Create(Vec v) noexcept
842: {
843:   PetscBool          alloc_missing;
844:   PetscDeviceContext dctx;

846:   PetscFunctionBegin;
847:   PetscCall(VecCreate_IMPL_Private(v, &alloc_missing));
848:   PetscCall(GetHandles_(&dctx));
849:   PetscCall(Initialize_CUPMBase(v, alloc_missing, nullptr, nullptr, dctx));
850:   PetscFunctionReturn(PETSC_SUCCESS);
851: }

853: // v->ops->destroy
854: template <device::cupm::DeviceType T, typename D>
855: inline PetscErrorCode Vec_CUPMBase<T, D>::Destroy(Vec v) noexcept
856: {
857:   PetscFunctionBegin;
858:   PetscCall(DestroyDevice_(v));
859:   PetscCall(DestroyHost_(v));
860:   PetscFunctionReturn(PETSC_SUCCESS);
861: }

863: // ================================================================================== //
864: //                      Common core between Seq and MPI                               //

866: // VecCreate_CUPM()
867: template <device::cupm::DeviceType T, typename D>
868: inline PetscErrorCode Vec_CUPMBase<T, D>::Create_CUPM(Vec v) noexcept
869: {
870:   PetscMPIInt size;

872:   PetscFunctionBegin;
873:   PetscCallMPI(MPI_Comm_size(PetscObjectComm(PetscObjectCast(v)), &size));
874:   PetscCall(VecSetType(v, size > 1 ? VECMPICUPM() : VECSEQCUPM()));
875:   PetscFunctionReturn(PETSC_SUCCESS);
876: }

878: // VecCreateCUPM()
879: template <device::cupm::DeviceType T, typename D>
880: inline PetscErrorCode Vec_CUPMBase<T, D>::Create_CUPMBase(MPI_Comm comm, PetscInt bs, PetscInt n, PetscInt N, Vec *v, PetscBool call_set_type, PetscLayout reference) noexcept
881: {
882:   PetscFunctionBegin;
883:   PetscCall(VecCreate(comm, v));
884:   if (reference) PetscCall(PetscLayoutReference(reference, &(*v)->map));
885:   PetscCall(VecSetSizes(*v, n, N));
886:   if (bs) PetscCall(VecSetBlockSize(*v, bs));
887:   if (call_set_type) PetscCall(VecSetType(*v, VECIMPLCUPM()));
888:   PetscFunctionReturn(PETSC_SUCCESS);
889: }

891: // VecCreateIMPL_CUPM(), called through v->ops->create
892: template <device::cupm::DeviceType T, typename D>
893: inline PetscErrorCode Vec_CUPMBase<T, D>::Initialize_CUPMBase(Vec v, PetscBool allocate_missing, PetscScalar *host_array, PetscScalar *device_array, PetscDeviceContext dctx) noexcept
894: {
895:   PetscFunctionBegin;
896:   // REVIEW ME: perhaps not needed
897:   PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUPM()));
898:   PetscCall(PetscObjectChangeTypeName(PetscObjectCast(v), VECIMPLCUPM()));
899:   PetscCall(D::BindToCPU(v, PETSC_FALSE));
900:   if (device_array) {
901:     PetscCall(CheckPointerMatchesMemType_(device_array, PETSC_MEMTYPE_CUPM()));
902:     PetscCall(VecCUPMAllocateCheck_(v));
903:     VecCUPMCast(v)->array_d = device_array;
904:   }
905:   if (host_array) {
906:     PetscCall(CheckPointerMatchesMemType_(host_array, PETSC_MEMTYPE_HOST));
907:     VecIMPLCast(v)->array = host_array;
908:   }
909:   if (allocate_missing) {
910:     PetscCall(DeviceAllocateCheck_(dctx, v));
911:     PetscCall(HostAllocateCheck_(dctx, v));
912:     // REVIEW ME: junchao, is this needed with new calloc() branch? VecSet() will call
913:     // set() for reference
914:     // calls device-version
915:     PetscCall(VecSet(v, 0));
916:     // zero the host while device is underway
917:     PetscCall(PetscArrayzero(VecIMPLCast(v)->array, v->map->n));
918:     v->offloadmask = PETSC_OFFLOAD_BOTH;
919:   } else {
920:     if (host_array) {
921:       v->offloadmask = device_array ? PETSC_OFFLOAD_BOTH : PETSC_OFFLOAD_CPU;
922:     } else {
923:       v->offloadmask = device_array ? PETSC_OFFLOAD_GPU : PETSC_OFFLOAD_UNALLOCATED;
924:     }
925:   }
926:   PetscFunctionReturn(PETSC_SUCCESS);
927: }

929: // v->ops->duplicate
930: template <device::cupm::DeviceType T, typename D>
931: template <typename SetupFunctionT>
932: inline PetscErrorCode Vec_CUPMBase<T, D>::Duplicate_CUPMBase(Vec v, Vec *y, PetscDeviceContext dctx, SetupFunctionT &&DerivedCreateIMPLCUPM_Async) noexcept
933: {
934:   // if the derived setup is the default no_op then we should call VecSetType()
935:   constexpr auto call_set_type = static_cast<PetscBool>(std::is_same<SetupFunctionT, no_op>::value);
936:   const auto     vobj          = PetscObjectCast(v);
937:   const auto     map           = v->map;
938:   PetscInt       bs;

940:   PetscFunctionBegin;
941:   PetscCall(VecGetBlockSize(v, &bs));
942:   PetscCall(Create_CUPMBase(PetscObjectComm(vobj), bs, map->n, map->N, y, call_set_type, map));
943:   // Derived class can set up the remainder of the data structures here
944:   PetscCall(DerivedCreateIMPLCUPM_Async(*y));
945:   // If the other vector is bound to CPU then the memcpy of the ops struct will give the
946:   // duplicated vector the host "getarray" function which does not lazily allocate the array
947:   // (as it is assumed to always exist). So we force allocation here, before we overwrite the
948:   // ops
949:   if (v->boundtocpu) PetscCall(HostAllocateCheck_(dctx, *y));
950:   // in case the user has done some VecSetOps() tomfoolery
951:   (*y)->ops[0] = v->ops[0];
952:   {
953:     const auto yobj = PetscObjectCast(*y);

955:     PetscCall(PetscObjectListDuplicate(vobj->olist, &yobj->olist));
956:     PetscCall(PetscFunctionListDuplicate(vobj->qlist, &yobj->qlist));
957:   }
958:   (*y)->stash.donotstash   = v->stash.donotstash;
959:   (*y)->stash.ignorenegidx = v->stash.ignorenegidx;
960:   (*y)->map->bs            = std::abs(v->map->bs);
961:   (*y)->bstash.bs          = v->bstash.bs;
962:   PetscFunctionReturn(PETSC_SUCCESS);
963: }

965:   #define VecSetOp_CUPM(op_name, op_host, ...) \
966:     do { \
967:       if (usehost) { \
968:         v->ops->op_name = op_host; \
969:       } else { \
970:         v->ops->op_name = __VA_ARGS__; \
971:       } \
972:     } while (0)

974: // v->ops->bindtocpu
975: template <device::cupm::DeviceType T, typename D>
976: inline PetscErrorCode Vec_CUPMBase<T, D>::BindToCPU_CUPMBase(Vec v, PetscBool usehost, PetscDeviceContext dctx) noexcept
977: {
978:   PetscFunctionBegin;
979:   v->boundtocpu = usehost;
980:   if (usehost) PetscCall(CopyToHost_(dctx, v));
981:   PetscCall(PetscStrFreeAllocpy(usehost ? PETSCRANDER48 : PETSCDEVICERAND(), &v->defaultrandtype));

983:   // set the base functions that are guaranteed to be the same for both
984:   v->ops->duplicate = D::Duplicate;
985:   v->ops->create    = D::Create;
986:   v->ops->destroy   = D::Destroy;
987:   v->ops->bindtocpu = D::BindToCPU;
988:   // Note that setting these to NULL on host breaks convergence in certain areas. I don't know
989:   // why, and I don't know how, but it is IMPERATIVE these are set as such!
990:   v->ops->replacearray = D::template ReplaceArray<PETSC_MEMTYPE_HOST>;
991:   v->ops->restorearray = D::template RestoreArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ_WRITE>;

993:   // set device-only common functions
994:   VecSetOp_CUPM(getarray, nullptr, D::template GetArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ_WRITE>);
995:   VecSetOp_CUPM(getarraywrite, nullptr, D::template GetArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_WRITE>);
996:   VecSetOp_CUPM(restorearraywrite, nullptr, D::template RestoreArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_WRITE>);

998:   VecSetOp_CUPM(getarrayread, nullptr, [](Vec v, const PetscScalar **a) { return D::template GetArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ>(v, const_cast<PetscScalar **>(a)); });
999:   VecSetOp_CUPM(restorearrayread, nullptr, [](Vec v, const PetscScalar **a) { return D::template RestoreArray<PETSC_MEMTYPE_HOST, PETSC_MEMORY_ACCESS_READ>(v, const_cast<PetscScalar **>(a)); });

1001:   VecSetOp_CUPM(getarrayandmemtype, nullptr, D::template GetArrayAndMemtype<PETSC_MEMORY_ACCESS_READ_WRITE>);
1002:   VecSetOp_CUPM(restorearrayandmemtype, nullptr, D::template RestoreArrayAndMemtype<PETSC_MEMORY_ACCESS_READ_WRITE>);

1004:   VecSetOp_CUPM(getarraywriteandmemtype, nullptr, D::template GetArrayAndMemtype<PETSC_MEMORY_ACCESS_WRITE>);
1005:   VecSetOp_CUPM(restorearraywriteandmemtype, nullptr, [](Vec v, PetscScalar **a, PetscMemType *) { return D::template RestoreArrayAndMemtype<PETSC_MEMORY_ACCESS_WRITE>(v, a); });

1007:   VecSetOp_CUPM(getarrayreadandmemtype, nullptr, [](Vec v, const PetscScalar **a, PetscMemType *m) { return D::template GetArrayAndMemtype<PETSC_MEMORY_ACCESS_READ>(v, const_cast<PetscScalar **>(a), m); });
1008:   VecSetOp_CUPM(restorearrayreadandmemtype, nullptr, [](Vec v, const PetscScalar **a) { return D::template RestoreArrayAndMemtype<PETSC_MEMORY_ACCESS_READ>(v, const_cast<PetscScalar **>(a)); });

1010:   // set the functions that are always sequential
1011:   using VecSeq_T = VecSeq_CUPM<T>;
1012:   VecSetOp_CUPM(scale, VecScale_Seq, VecSeq_T::Scale);
1013:   VecSetOp_CUPM(copy, VecCopy_Seq, VecSeq_T::Copy);
1014:   VecSetOp_CUPM(set, VecSet_Seq, VecSeq_T::Set);
1015:   VecSetOp_CUPM(swap, VecSwap_Seq, VecSeq_T::Swap);
1016:   VecSetOp_CUPM(axpy, VecAXPY_Seq, VecSeq_T::AXPY);
1017:   VecSetOp_CUPM(axpby, VecAXPBY_Seq, VecSeq_T::AXPBY);
1018:   VecSetOp_CUPM(maxpy, VecMAXPY_Seq, VecSeq_T::MAXPY);
1019:   VecSetOp_CUPM(aypx, VecAYPX_Seq, VecSeq_T::AYPX);
1020:   VecSetOp_CUPM(waxpy, VecWAXPY_Seq, VecSeq_T::WAXPY);
1021:   VecSetOp_CUPM(axpbypcz, VecAXPBYPCZ_Seq, VecSeq_T::AXPBYPCZ);
1022:   VecSetOp_CUPM(pointwisemult, VecPointwiseMult_Seq, VecSeq_T::PointwiseMult);
1023:   VecSetOp_CUPM(pointwisedivide, VecPointwiseDivide_Seq, VecSeq_T::PointwiseDivide);
1024:   VecSetOp_CUPM(pointwisemax, VecPointwiseMax_Seq, VecSeq_T::PointwiseMax);
1025:   VecSetOp_CUPM(pointwisemaxabs, VecPointwiseMaxAbs_Seq, VecSeq_T::PointwiseMaxAbs);
1026:   VecSetOp_CUPM(pointwisemin, VecPointwiseMin_Seq, VecSeq_T::PointwiseMin);
1027:   VecSetOp_CUPM(setrandom, VecSetRandom_Seq, VecSeq_T::SetRandom);
1028:   VecSetOp_CUPM(dot_local, VecDot_Seq, VecSeq_T::Dot);
1029:   VecSetOp_CUPM(tdot_local, VecTDot_Seq, VecSeq_T::TDot);
1030:   VecSetOp_CUPM(norm_local, VecNorm_Seq, VecSeq_T::Norm);
1031:   VecSetOp_CUPM(mdot_local, VecMDot_Seq, VecSeq_T::MDot);
1032:   VecSetOp_CUPM(reciprocal, VecReciprocal_Default, VecSeq_T::Reciprocal);
1033:   VecSetOp_CUPM(conjugate, VecConjugate_Seq, VecSeq_T::Conjugate);
1034:   VecSetOp_CUPM(abs, nullptr, VecSeq_T::Abs);
1035:   VecSetOp_CUPM(sqrt, nullptr, VecSeq_T::SqrtAbs);
1036:   VecSetOp_CUPM(exp, nullptr, VecSeq_T::Exp);
1037:   VecSetOp_CUPM(log, nullptr, VecSeq_T::Log);
1038:   VecSetOp_CUPM(shift, nullptr, VecSeq_T::Shift);
1039:   VecSetOp_CUPM(dotnorm2, nullptr, D::DotNorm2);
1040:   VecSetOp_CUPM(getlocalvector, nullptr, VecSeq_T::template GetLocalVector<PETSC_MEMORY_ACCESS_READ_WRITE>);
1041:   VecSetOp_CUPM(restorelocalvector, nullptr, VecSeq_T::template RestoreLocalVector<PETSC_MEMORY_ACCESS_READ_WRITE>);
1042:   VecSetOp_CUPM(getlocalvectorread, nullptr, VecSeq_T::template GetLocalVector<PETSC_MEMORY_ACCESS_READ>);
1043:   VecSetOp_CUPM(restorelocalvectorread, nullptr, VecSeq_T::template RestoreLocalVector<PETSC_MEMORY_ACCESS_READ>);
1044:   VecSetOp_CUPM(sum, nullptr, VecSeq_T::Sum);
1045:   VecSetOp_CUPM(errorwnorm, nullptr, D::ErrorWnorm);
1046:   VecSetOp_CUPM(duplicatevecs, VecDuplicateVecs_Default, VecDuplicateVecs_Default);
1047:   VecSetOp_CUPM(setstdbasis, nullptr, VecSeq_T::SetStdBasis);
1048:   PetscFunctionReturn(PETSC_SUCCESS);
1049: }

1051: // Called from VecGetSubVector()
1052: template <device::cupm::DeviceType T, typename D>
1053: inline PetscErrorCode Vec_CUPMBase<T, D>::GetArrays_CUPMBase(Vec v, const PetscScalar **host_array, const PetscScalar **device_array, PetscOffloadMask *mask, PetscDeviceContext dctx) noexcept
1054: {
1055:   PetscFunctionBegin;
1056:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
1057:   if (host_array) {
1058:     PetscCall(HostAllocateCheck_(dctx, v));
1059:     *host_array = VecIMPLCast(v)->array;
1060:   }
1061:   if (device_array) {
1062:     PetscCall(DeviceAllocateCheck_(dctx, v));
1063:     *device_array = VecCUPMCast(v)->array_d;
1064:   }
1065:   if (mask) *mask = v->offloadmask;
1066:   PetscFunctionReturn(PETSC_SUCCESS);
1067: }

1069: template <device::cupm::DeviceType T, typename D>
1070: inline PetscErrorCode Vec_CUPMBase<T, D>::ResetPreallocationCOO_CUPMBase(Vec v, PetscDeviceContext dctx) noexcept
1071: {
1072:   PetscFunctionBegin;
1073:   if (const auto vcu = VecCUPMCast(v)) {
1074:     cupmStream_t stream;
1075:     // clang-format off
1076:     const auto   cntptrs = util::make_array(
1077:       std::ref(vcu->jmap1_d),
1078:       std::ref(vcu->perm1_d),
1079:       std::ref(vcu->imap2_d),
1080:       std::ref(vcu->jmap2_d),
1081:       std::ref(vcu->perm2_d),
1082:       std::ref(vcu->Cperm_d)
1083:     );
1084:     // clang-format on

1086:     PetscCall(GetHandlesFrom_(dctx, &stream));
1087:     for (auto &&ptr : cntptrs) PetscCallCUPM(cupmFreeAsync(ptr.get(), stream));
1088:     for (auto &&ptr : util::make_array(std::ref(vcu->sendbuf_d), std::ref(vcu->recvbuf_d))) PetscCallCUPM(cupmFreeAsync(ptr.get(), stream));
1089:   }
1090:   PetscFunctionReturn(PETSC_SUCCESS);
1091: }

1093: template <device::cupm::DeviceType T, typename D>
1094: template <std::size_t NCount, std::size_t NScal>
1095: inline PetscErrorCode Vec_CUPMBase<T, D>::SetPreallocationCOO_CUPMBase(Vec v, PetscCount, const PetscInt[], PetscDeviceContext dctx, const std::array<CooPair<PetscCount>, NCount> &extra_cntptrs, const std::array<CooPair<PetscScalar>, NScal> &bufptrs) noexcept
1096: {
1097:   PetscFunctionBegin;
1098:   PetscCall(ResetPreallocationCOO_CUPMBase(v, dctx));
1099:   // need to instantiate the private pointer if not already
1100:   PetscCall(VecCUPMAllocateCheck_(v));
1101:   {
1102:     const auto vimpl = VecIMPLCast(v);
1103:     const auto vcu   = VecCUPMCast(v);
1104:     // clang-format off
1105:     const auto cntptrs = util::concat_array(
1106:       util::make_array(
1107:         make_coo_pair(vcu->jmap1_d, vimpl->jmap1, v->map->n + 1),
1108:         make_coo_pair(vcu->perm1_d, vimpl->perm1, vimpl->tot1)
1109:       ),
1110:       extra_cntptrs
1111:     );
1112:     // clang-format on
1113:     cupmStream_t stream;

1115:     PetscCall(GetHandlesFrom_(dctx, &stream));
1116:     // allocate
1117:     for (auto &elem : cntptrs) PetscCall(PetscCUPMMallocAsync(&elem.device, elem.size, stream));
1118:     for (auto &elem : bufptrs) PetscCall(PetscCUPMMallocAsync(&elem.device, elem.size, stream));
1119:     // copy
1120:     for (const auto &elem : cntptrs) PetscCall(PetscCUPMMemcpyAsync(elem.device, elem.host, elem.size, cupmMemcpyHostToDevice, stream, true));
1121:     for (const auto &elem : bufptrs) PetscCall(PetscCUPMMemcpyAsync(elem.device, elem.host, elem.size, cupmMemcpyHostToDevice, stream, true));
1122:   }
1123:   PetscFunctionReturn(PETSC_SUCCESS);
1124: }

1126: template <device::cupm::DeviceType T, typename D>
1127: inline PetscErrorCode Vec_CUPMBase<T, D>::Convert_IMPL_IMPLCUPM(Vec v) noexcept
1128: {
1129:   const auto         n        = v->map->n;
1130:   const auto         vimpl    = VecIMPLCast(v);
1131:   auto              &impl_arr = vimpl->array;
1132:   PetscBool          set      = PETSC_FALSE;
1133:   PetscDeviceContext dctx;

1135:   PetscFunctionBegin;
1136:   // If users do not explicitly require pinned memory, we prefer keeping the vector's regular
1137:   // host array
1138:   PetscCall(VecCUPMCheckMinimumPinnedMemory_Internal(v, &set));
1139:   if (set && impl_arr && ((n * sizeof(*impl_arr)) > v->minimum_bytes_pinned_memory)) {
1140:     auto        &impl_alloc = vimpl->array_allocated;
1141:     PetscScalar *new_arr;

1143:     // users require pinned memory
1144:     {
1145:       // Allocate pinned memory and copy over the old array
1146:       const auto useit = UseCUPMHostAlloc(PETSC_TRUE);

1148:       PetscCall(PetscMalloc1(n, &new_arr));
1149:       PetscCall(PetscArraycpy(new_arr, impl_arr, n));
1150:     }
1151:     PetscCall(PetscFree(impl_alloc));
1152:     impl_arr         = new_arr;
1153:     impl_alloc       = new_arr;
1154:     v->offloadmask   = PETSC_OFFLOAD_CPU;
1155:     v->pinned_memory = PETSC_TRUE;
1156:   }
1157:   PetscCall(GetHandles_(&dctx));
1158:   PetscCall(Initialize_CUPMBase(v, PETSC_FALSE, impl_arr, nullptr, dctx));
1159:   PetscFunctionReturn(PETSC_SUCCESS);
1160: }

1162:   #define PETSC_VEC_CUPM_BASE_CLASS_HEADER(name, Tp, ...) \
1163:     PETSC_CUPMOBJECT_HEADER(Tp); \
1164:     using name = ::Petsc::vec::cupm::impl::Vec_CUPMBase<Tp, __VA_ARGS__>; \
1165:     friend name; \
1166:     /* introspection */ \
1167:     using name::VecCUPMCast; \
1168:     using name::VecIMPLCast; \
1169:     using name::VECIMPLCUPM; \
1170:     using name::VECIMPL; \
1171:     using name::VECSEQCUPM; \
1172:     using name::VECMPICUPM; \
1173:     using name::VECCUPM; \
1174:     using name::VecView_Debug; \
1175:     /* utility */ \
1176:     using typename name::Vec_CUPM; \
1177:     using name::VecCUPMAllocateCheck_; \
1178:     using name::VecIMPLAllocateCheck_; \
1179:     using name::HostAllocateCheck_; \
1180:     using name::DeviceAllocateCheck_; \
1181:     using name::CopyToDevice_; \
1182:     using name::CopyToHost_; \
1183:     using name::Create; \
1184:     using name::Destroy; \
1185:     using name::GetArray; \
1186:     using name::RestoreArray; \
1187:     using name::GetArrayAndMemtype; \
1188:     using name::RestoreArrayAndMemtype; \
1189:     using name::PlaceArray; \
1190:     using name::ReplaceArray; \
1191:     using name::ResetArray; \
1192:     /* base functions */ \
1193:     using name::Create_CUPMBase; \
1194:     using name::Initialize_CUPMBase; \
1195:     using name::Duplicate_CUPMBase; \
1196:     using name::BindToCPU_CUPMBase; \
1197:     using name::Create_CUPM; \
1198:     using name::DeviceArrayRead; \
1199:     using name::DeviceArrayWrite; \
1200:     using name::DeviceArrayReadWrite; \
1201:     using name::HostArrayRead; \
1202:     using name::HostArrayWrite; \
1203:     using name::HostArrayReadWrite; \
1204:     using name::ResetPreallocationCOO_CUPMBase; \
1205:     using name::SetPreallocationCOO_CUPMBase; \
1206:     using name::Convert_IMPL_IMPLCUPM;

1208: } // namespace impl

1210: } // namespace cupm

1212: } // namespace vec

1214: } // namespace Petsc

1216: #endif // __cplusplus && PetscDefined(HAVE_DEVICE)