Actual source code: vecseqcupm_impl.hpp

  1: #pragma once

  3: #include "vecseqcupm.hpp"

  5: #include <petsc/private/randomimpl.h>

  7: #include "../src/sys/objects/device/impls/cupm/cupmthrustutility.hpp"
  8: #include "../src/sys/objects/device/impls/cupm/kernels.hpp"

 10: #if PetscDefined(USE_COMPLEX)
 11:   #include <thrust/transform_reduce.h>
 12: #endif
 13: #include <thrust/transform.h>
 14: #include <thrust/reduce.h>
 15: #include <thrust/functional.h>
 16: #include <thrust/tuple.h>
 17: #include <thrust/device_ptr.h>
 18: #include <thrust/iterator/zip_iterator.h>
 19: #include <thrust/iterator/counting_iterator.h>
 20: #include <thrust/iterator/constant_iterator.h>
 21: #include <thrust/inner_product.h>

 23: namespace Petsc
 24: {

 26: namespace vec
 27: {

 29: namespace cupm
 30: {

 32: namespace impl
 33: {

 35: // ==========================================================================================
 36: // VecSeq_CUPM - Private API
 37: // ==========================================================================================

 39: template <device::cupm::DeviceType T>
 40: inline Vec_Seq *VecSeq_CUPM<T>::VecIMPLCast_(Vec v) noexcept
 41: {
 42:   return static_cast<Vec_Seq *>(v->data);
 43: }

 45: template <device::cupm::DeviceType T>
 46: inline constexpr VecType VecSeq_CUPM<T>::VECIMPLCUPM_() noexcept
 47: {
 48:   return VECSEQCUPM();
 49: }

 51: template <device::cupm::DeviceType T>
 52: inline constexpr VecType VecSeq_CUPM<T>::VECIMPL_() noexcept
 53: {
 54:   return VECSEQ;
 55: }

 57: template <device::cupm::DeviceType T>
 58: inline PetscErrorCode VecSeq_CUPM<T>::ClearAsyncFunctions(Vec v) noexcept
 59: {
 60:   PetscFunctionBegin;
 61:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Abs), nullptr));
 62:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPBY), nullptr));
 63:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPBYPCZ), nullptr));
 64:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPY), nullptr));
 65:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AYPX), nullptr));
 66:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Conjugate), nullptr));
 67:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Copy), nullptr));
 68:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Exp), nullptr));
 69:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Log), nullptr));
 70:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(MAXPY), nullptr));
 71:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseDivide), nullptr));
 72:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMax), nullptr));
 73:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMaxAbs), nullptr));
 74:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMin), nullptr));
 75:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMult), nullptr));
 76:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Reciprocal), nullptr));
 77:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Scale), nullptr));
 78:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Set), nullptr));
 79:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Shift), nullptr));
 80:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(SqrtAbs), nullptr));
 81:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Swap), nullptr));
 82:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(WAXPY), nullptr));
 83:   PetscFunctionReturn(PETSC_SUCCESS);
 84: }

 86: template <device::cupm::DeviceType T>
 87: inline PetscErrorCode VecSeq_CUPM<T>::InitializeAsyncFunctions(Vec v) noexcept
 88: {
 89:   PetscFunctionBegin;
 90:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Abs), VecSeq_CUPM<T>::AbsAsync));
 91:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPBY), VecSeq_CUPM<T>::AXPBYAsync));
 92:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPBYPCZ), VecSeq_CUPM<T>::AXPBYPCZAsync));
 93:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AXPY), VecSeq_CUPM<T>::AXPYAsync));
 94:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(AYPX), VecSeq_CUPM<T>::AYPXAsync));
 95:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Conjugate), VecSeq_CUPM<T>::ConjugateAsync));
 96:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Copy), VecSeq_CUPM<T>::CopyAsync));
 97:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Exp), VecSeq_CUPM<T>::ExpAsync));
 98:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Log), VecSeq_CUPM<T>::LogAsync));
 99:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(MAXPY), VecSeq_CUPM<T>::MAXPYAsync));
100:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseDivide), VecSeq_CUPM<T>::PointwiseDivideAsync));
101:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMax), VecSeq_CUPM<T>::PointwiseMaxAsync));
102:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMaxAbs), VecSeq_CUPM<T>::PointwiseMaxAbsAsync));
103:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMin), VecSeq_CUPM<T>::PointwiseMinAsync));
104:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(PointwiseMult), VecSeq_CUPM<T>::PointwiseMultAsync));
105:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Reciprocal), VecSeq_CUPM<T>::ReciprocalAsync));
106:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Scale), VecSeq_CUPM<T>::ScaleAsync));
107:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Set), VecSeq_CUPM<T>::SetAsync));
108:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Shift), VecSeq_CUPM<T>::ShiftAsync));
109:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(SqrtAbs), VecSeq_CUPM<T>::SqrtAbsAsync));
110:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(Swap), VecSeq_CUPM<T>::SwapAsync));
111:   PetscCall(PetscObjectComposeFunction(PetscObjectCast(v), VecAsyncFnName(WAXPY), VecSeq_CUPM<T>::WAXPYAsync));
112:   PetscFunctionReturn(PETSC_SUCCESS);
113: }

115: template <device::cupm::DeviceType T>
116: inline PetscErrorCode VecSeq_CUPM<T>::VecDestroy_IMPL_(Vec v) noexcept
117: {
118:   PetscFunctionBegin;
119:   PetscCall(ClearAsyncFunctions(v));
120:   PetscCall(VecDestroy_Seq(v));
121:   PetscFunctionReturn(PETSC_SUCCESS);
122: }

124: template <device::cupm::DeviceType T>
125: inline PetscErrorCode VecSeq_CUPM<T>::VecResetArray_IMPL_(Vec v) noexcept
126: {
127:   return VecResetArray_Seq(v);
128: }

130: template <device::cupm::DeviceType T>
131: inline PetscErrorCode VecSeq_CUPM<T>::VecPlaceArray_IMPL_(Vec v, const PetscScalar *a) noexcept
132: {
133:   return VecPlaceArray_Seq(v, a);
134: }

136: template <device::cupm::DeviceType T>
137: inline PetscErrorCode VecSeq_CUPM<T>::VecCreate_IMPL_Private_(Vec v, PetscBool *alloc_missing, PetscInt, PetscScalar *host_array) noexcept
138: {
139:   PetscMPIInt size;

141:   PetscFunctionBegin;
142:   if (alloc_missing) *alloc_missing = PETSC_FALSE;
143:   PetscCallMPI(MPI_Comm_size(PetscObjectComm(PetscObjectCast(v)), &size));
144:   PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Must create VecSeq on communicator of size 1, have size %d", size);
145:   PetscCall(VecCreate_Seq_Private(v, host_array));
146:   PetscCall(InitializeAsyncFunctions(v));
147:   PetscFunctionReturn(PETSC_SUCCESS);
148: }

150: // for functions with an early return based one vec size we still need to artificially bump the
151: // object state. This is to prevent the following:
152: //
153: // 0. Suppose you have a Vec {
154: //   rank 0: [0],
155: //   rank 1: [<empty>]
156: // }
157: // 1. both ranks have Vec with PetscObjectState = 0, stashed norm of 0
158: // 2. Vec enters e.g. VecSet(10)
159: // 3. rank 1 has local size 0 and bails immediately
160: // 4. rank 0 has local size 1 and enters function, eventually calls DeviceArrayWrite()
161: // 5. DeviceArrayWrite() calls PetscObjectStateIncrease(), now state = 1
162: // 6. Vec enters VecNorm(), and calls VecNormAvailable()
163: // 7. rank 1 has object state = 0, equal to stash and returns early with norm = 0
164: // 8. rank 0 has object state = 1, not equal to stash, continues to impl function
165: // 9. rank 0 deadlocks on MPI_Allreduce() because rank 1 bailed early
166: template <device::cupm::DeviceType T>
167: inline PetscErrorCode VecSeq_CUPM<T>::MaybeIncrementEmptyLocalVec(Vec v) noexcept
168: {
169:   PetscFunctionBegin;
170:   if (PetscUnlikely((v->map->n == 0) && (v->map->N != 0))) PetscCall(PetscObjectStateIncrease(PetscObjectCast(v)));
171:   PetscFunctionReturn(PETSC_SUCCESS);
172: }

174: template <device::cupm::DeviceType T>
175: inline PetscErrorCode VecSeq_CUPM<T>::CreateSeqCUPM_(Vec v, PetscDeviceContext dctx, PetscScalar *host_array, PetscScalar *device_array) noexcept
176: {
177:   PetscFunctionBegin;
178:   PetscCall(base_type::VecCreate_IMPL_Private(v, nullptr, 0, host_array));
179:   PetscCall(Initialize_CUPMBase(v, PETSC_FALSE, host_array, device_array, dctx));
180:   PetscFunctionReturn(PETSC_SUCCESS);
181: }

183: template <device::cupm::DeviceType T>
184: template <typename BinaryFuncT>
185: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseBinary_(BinaryFuncT &&binary, Vec xin, Vec yin, Vec zout, PetscDeviceContext dctx) noexcept
186: {
187:   PetscFunctionBegin;
188:   if (const auto n = zout->map->n) {
189:     cupmStream_t stream;

191:     PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
192:     PetscCall(GetHandlesFrom_(dctx, &stream));
193:     // clang-format off
194:     PetscCallThrust(
195:       const auto dxptr = thrust::device_pointer_cast(DeviceArrayRead(dctx, xin).data());

197:       THRUST_CALL(
198:         thrust::transform,
199:         stream,
200:         dxptr, dxptr + n,
201:         thrust::device_pointer_cast(DeviceArrayRead(dctx, yin).data()),
202:         thrust::device_pointer_cast(DeviceArrayWrite(dctx, zout).data()),
203:         std::forward<BinaryFuncT>(binary)
204:       )
205:     );
206:     // clang-format on
207:     PetscCall(PetscLogGpuFlops(n));
208:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
209:   } else {
210:     PetscCall(MaybeIncrementEmptyLocalVec(zout));
211:   }
212:   PetscFunctionReturn(PETSC_SUCCESS);
213: }

215: template <device::cupm::DeviceType T>
216: template <typename BinaryFuncT>
217: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseBinaryDispatch_(PetscErrorCode (*VecSeqFunction)(Vec, Vec, Vec), BinaryFuncT &&binary, Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
218: {
219:   PetscFunctionBegin;
220:   if (xin->boundtocpu || yin->boundtocpu) {
221:     PetscCall((*VecSeqFunction)(wout, xin, yin));
222:   } else {
223:     // note order of arguments! xin and yin are read, wout is written!
224:     PetscCall(PointwiseBinary_(std::forward<BinaryFuncT>(binary), xin, yin, wout, dctx));
225:   }
226:   PetscFunctionReturn(PETSC_SUCCESS);
227: }

229: template <device::cupm::DeviceType T>
230: template <typename UnaryFuncT>
231: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseUnary_(UnaryFuncT &&unary, Vec xinout, Vec yin, PetscDeviceContext dctx) noexcept
232: {
233:   const auto inplace = !yin || (xinout == yin);

235:   PetscFunctionBegin;
236:   if (const auto n = xinout->map->n) {
237:     cupmStream_t stream;
238:     const auto   apply = [&](PetscScalar *xinout, PetscScalar *yin = nullptr) {
239:       PetscFunctionBegin;
240:       // clang-format off
241:       PetscCallThrust(
242:         const auto xptr = thrust::device_pointer_cast(xinout);

244:         THRUST_CALL(
245:           thrust::transform,
246:           stream,
247:           xptr, xptr + n,
248:           (yin && (yin != xinout)) ? thrust::device_pointer_cast(yin) : xptr,
249:           std::forward<UnaryFuncT>(unary)
250:         )
251:       );
252:       // clang-format on
253:       PetscFunctionReturn(PETSC_SUCCESS);
254:     };

256:     PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
257:     PetscCall(GetHandlesFrom_(dctx, &stream));
258:     if (inplace) {
259:       PetscCall(apply(DeviceArrayReadWrite(dctx, xinout).data()));
260:     } else {
261:       PetscCall(apply(DeviceArrayRead(dctx, xinout).data(), DeviceArrayWrite(dctx, yin).data()));
262:     }
263:     PetscCall(PetscLogGpuFlops(n));
264:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
265:   } else {
266:     if (inplace) {
267:       PetscCall(MaybeIncrementEmptyLocalVec(xinout));
268:     } else {
269:       PetscCall(MaybeIncrementEmptyLocalVec(yin));
270:     }
271:   }
272:   PetscFunctionReturn(PETSC_SUCCESS);
273: }

275: // ==========================================================================================
276: // VecSeq_CUPM - Public API - Constructors
277: // ==========================================================================================

279: // VecCreateSeqCUPM()
280: template <device::cupm::DeviceType T>
281: inline PetscErrorCode VecSeq_CUPM<T>::CreateSeqCUPM(MPI_Comm comm, PetscInt bs, PetscInt n, Vec *v, PetscBool call_set_type) noexcept
282: {
283:   PetscFunctionBegin;
284:   PetscCall(Create_CUPMBase(comm, bs, n, n, v, call_set_type));
285:   PetscFunctionReturn(PETSC_SUCCESS);
286: }

288: // VecCreateSeqCUPMWithArrays()
289: template <device::cupm::DeviceType T>
290: inline PetscErrorCode VecSeq_CUPM<T>::CreateSeqCUPMWithBothArrays(MPI_Comm comm, PetscInt bs, PetscInt n, const PetscScalar host_array[], const PetscScalar device_array[], Vec *v) noexcept
291: {
292:   PetscDeviceContext dctx;

294:   PetscFunctionBegin;
295:   PetscCall(GetHandles_(&dctx));
296:   // do NOT call VecSetType(), otherwise ops->create() -> create() ->
297:   // CreateSeqCUPM_() is called!
298:   PetscCall(CreateSeqCUPM(comm, bs, n, v, PETSC_FALSE));
299:   PetscCall(CreateSeqCUPM_(*v, dctx, PetscRemoveConstCast(host_array), PetscRemoveConstCast(device_array)));
300:   PetscFunctionReturn(PETSC_SUCCESS);
301: }

303: // v->ops->duplicate
304: template <device::cupm::DeviceType T>
305: inline PetscErrorCode VecSeq_CUPM<T>::Duplicate(Vec v, Vec *y) noexcept
306: {
307:   PetscDeviceContext dctx;

309:   PetscFunctionBegin;
310:   PetscCall(GetHandles_(&dctx));
311:   PetscCall(Duplicate_CUPMBase(v, y, dctx));
312:   PetscFunctionReturn(PETSC_SUCCESS);
313: }

315: // ==========================================================================================
316: // VecSeq_CUPM - Public API - Utility
317: // ==========================================================================================

319: // v->ops->bindtocpu
320: template <device::cupm::DeviceType T>
321: inline PetscErrorCode VecSeq_CUPM<T>::BindToCPU(Vec v, PetscBool usehost) noexcept
322: {
323:   PetscDeviceContext dctx;

325:   PetscFunctionBegin;
326:   PetscCall(GetHandles_(&dctx));
327:   PetscCall(BindToCPU_CUPMBase(v, usehost, dctx));

329:   // REVIEW ME: this absolutely should be some sort of bulk mempcy rather than this mess
330:   VecSetOp_CUPM(dot, VecDot_Seq, Dot);
331:   VecSetOp_CUPM(norm, VecNorm_Seq, Norm);
332:   VecSetOp_CUPM(tdot, VecTDot_Seq, TDot);
333:   VecSetOp_CUPM(mdot, VecMDot_Seq, MDot);
334:   VecSetOp_CUPM(resetarray, VecResetArray_Seq, base_type::template ResetArray<PETSC_MEMTYPE_HOST>);
335:   VecSetOp_CUPM(placearray, VecPlaceArray_Seq, base_type::template PlaceArray<PETSC_MEMTYPE_HOST>);
336:   v->ops->mtdot = v->ops->mtdot_local = VecMTDot_Seq;
337:   VecSetOp_CUPM(max, VecMax_Seq, Max);
338:   VecSetOp_CUPM(min, VecMin_Seq, Min);
339:   VecSetOp_CUPM(setpreallocationcoo, VecSetPreallocationCOO_Seq, SetPreallocationCOO);
340:   VecSetOp_CUPM(setvaluescoo, VecSetValuesCOO_Seq, SetValuesCOO);
341:   PetscFunctionReturn(PETSC_SUCCESS);
342: }

344: // ==========================================================================================
345: // VecSeq_CUPM - Public API - Mutators
346: // ==========================================================================================

348: // v->ops->getlocalvector or v->ops->getlocalvectorread
349: template <device::cupm::DeviceType T>
350: template <PetscMemoryAccessMode access>
351: inline PetscErrorCode VecSeq_CUPM<T>::GetLocalVector(Vec v, Vec w) noexcept
352: {
353:   PetscBool wisseqcupm;

355:   PetscFunctionBegin;
356:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
357:   PetscCall(PetscObjectTypeCompare(PetscObjectCast(w), VECSEQCUPM(), &wisseqcupm));
358:   if (wisseqcupm) {
359:     if (const auto wseq = VecIMPLCast(w)) {
360:       if (auto &alloced = wseq->array_allocated) {
361:         const auto useit = UseCUPMHostAlloc(util::exchange(w->pinned_memory, PETSC_FALSE));

363:         PetscCall(PetscFree(alloced));
364:       }
365:       wseq->array         = nullptr;
366:       wseq->unplacedarray = nullptr;
367:     }
368:     if (const auto wcu = VecCUPMCast(w)) {
369:       if (auto &device_array = wcu->array_d) {
370:         cupmStream_t stream;

372:         PetscCall(GetHandles_(&stream));
373:         PetscCallCUPM(cupmFreeAsync(device_array, stream));
374:       }
375:       PetscCall(PetscFree(w->spptr /* wcu */));
376:     }
377:   }
378:   if (v->petscnative && wisseqcupm) {
379:     PetscCall(PetscFree(w->data));
380:     w->data          = v->data;
381:     w->offloadmask   = v->offloadmask;
382:     w->pinned_memory = v->pinned_memory;
383:     w->spptr         = v->spptr;
384:     PetscCall(PetscObjectStateIncrease(PetscObjectCast(w)));
385:   } else {
386:     const auto array = &VecIMPLCast(w)->array;

388:     if (access == PETSC_MEMORY_ACCESS_READ) {
389:       PetscCall(VecGetArrayRead(v, const_cast<const PetscScalar **>(array)));
390:     } else {
391:       PetscCall(VecGetArray(v, array));
392:     }
393:     w->offloadmask = PETSC_OFFLOAD_CPU;
394:     if (wisseqcupm) {
395:       PetscDeviceContext dctx;

397:       PetscCall(GetHandles_(&dctx));
398:       PetscCall(DeviceAllocateCheck_(dctx, w));
399:     }
400:   }
401:   PetscFunctionReturn(PETSC_SUCCESS);
402: }

404: // v->ops->restorelocalvector or v->ops->restorelocalvectorread
405: template <device::cupm::DeviceType T>
406: template <PetscMemoryAccessMode access>
407: inline PetscErrorCode VecSeq_CUPM<T>::RestoreLocalVector(Vec v, Vec w) noexcept
408: {
409:   PetscBool wisseqcupm;

411:   PetscFunctionBegin;
412:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
413:   PetscCall(PetscObjectTypeCompare(PetscObjectCast(w), VECSEQCUPM(), &wisseqcupm));
414:   if (v->petscnative && wisseqcupm) {
415:     // the assignments to nullptr are __critical__, as w may persist after this call returns
416:     // and shouldn't share data with v!
417:     v->pinned_memory = w->pinned_memory;
418:     v->offloadmask   = util::exchange(w->offloadmask, PETSC_OFFLOAD_UNALLOCATED);
419:     v->data          = util::exchange(w->data, nullptr);
420:     v->spptr         = util::exchange(w->spptr, nullptr);
421:   } else {
422:     const auto array = &VecIMPLCast(w)->array;

424:     if (access == PETSC_MEMORY_ACCESS_READ) {
425:       PetscCall(VecRestoreArrayRead(v, const_cast<const PetscScalar **>(array)));
426:     } else {
427:       PetscCall(VecRestoreArray(v, array));
428:     }
429:     if (w->spptr && wisseqcupm) {
430:       cupmStream_t stream;

432:       PetscCall(GetHandles_(&stream));
433:       PetscCallCUPM(cupmFreeAsync(VecCUPMCast(w)->array_d, stream));
434:       PetscCall(PetscFree(w->spptr));
435:     }
436:   }
437:   PetscFunctionReturn(PETSC_SUCCESS);
438: }

440: // ==========================================================================================
441: // VecSeq_CUPM - Public API - Compute Methods
442: // ==========================================================================================

444: // VecAYPXAsync_Private
445: template <device::cupm::DeviceType T>
446: inline PetscErrorCode VecSeq_CUPM<T>::AYPXAsync(Vec yin, PetscScalar alpha, Vec xin, PetscDeviceContext dctx) noexcept
447: {
448:   const auto n = static_cast<cupmBlasInt_t>(yin->map->n);
449:   PetscBool  xiscupm;

451:   PetscFunctionBegin;
452:   PetscCall(PetscObjectTypeCompareAny(PetscObjectCast(xin), &xiscupm, VECSEQCUPM(), VECMPICUPM(), ""));
453:   if (!xiscupm) {
454:     PetscCall(VecAYPX_Seq(yin, alpha, xin));
455:     PetscFunctionReturn(PETSC_SUCCESS);
456:   }
457:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
458:   if (alpha == PetscScalar(0.0)) {
459:     cupmStream_t stream;

461:     PetscCall(GetHandlesFrom_(dctx, &stream));
462:     PetscCall(PetscLogGpuTimeBegin());
463:     PetscCall(PetscCUPMMemcpyAsync(DeviceArrayWrite(dctx, yin).data(), DeviceArrayRead(dctx, xin).data(), n, cupmMemcpyDeviceToDevice, stream));
464:     PetscCall(PetscLogGpuTimeEnd());
465:   } else if (n) {
466:     const auto       alphaIsOne = alpha == PetscScalar(1.0);
467:     const auto       calpha     = cupmScalarPtrCast(&alpha);
468:     cupmBlasHandle_t cupmBlasHandle;

470:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle));
471:     {
472:       const auto yptr = DeviceArrayReadWrite(dctx, yin);
473:       const auto xptr = DeviceArrayRead(dctx, xin);

475:       PetscCall(PetscLogGpuTimeBegin());
476:       if (alphaIsOne) {
477:         PetscCallCUPMBLAS(cupmBlasXaxpy(cupmBlasHandle, n, calpha, xptr.cupmdata(), 1, yptr.cupmdata(), 1));
478:       } else {
479:         const auto one = cupmScalarCast(1.0);

481:         PetscCallCUPMBLAS(cupmBlasXscal(cupmBlasHandle, n, calpha, yptr.cupmdata(), 1));
482:         PetscCallCUPMBLAS(cupmBlasXaxpy(cupmBlasHandle, n, &one, xptr.cupmdata(), 1, yptr.cupmdata(), 1));
483:       }
484:       PetscCall(PetscLogGpuTimeEnd());
485:     }
486:     PetscCall(PetscLogGpuFlops((alphaIsOne ? 1 : 2) * n));
487:   }
488:   if (n > 0) PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
489:   PetscFunctionReturn(PETSC_SUCCESS);
490: }

492: // v->ops->aypx
493: template <device::cupm::DeviceType T>
494: inline PetscErrorCode VecSeq_CUPM<T>::AYPX(Vec yin, PetscScalar alpha, Vec xin) noexcept
495: {
496:   PetscFunctionBegin;
497:   PetscCall(AYPXAsync(yin, alpha, xin, nullptr));
498:   PetscFunctionReturn(PETSC_SUCCESS);
499: }

501: // VecAXPYAsync_Private
502: template <device::cupm::DeviceType T>
503: inline PetscErrorCode VecSeq_CUPM<T>::AXPYAsync(Vec yin, PetscScalar alpha, Vec xin, PetscDeviceContext dctx) noexcept
504: {
505:   PetscBool xiscupm;

507:   PetscFunctionBegin;
508:   PetscCall(PetscObjectTypeCompareAny(PetscObjectCast(xin), &xiscupm, VECSEQCUPM(), VECMPICUPM(), ""));
509:   if (xiscupm) {
510:     const auto       n = static_cast<cupmBlasInt_t>(yin->map->n);
511:     cupmBlasHandle_t cupmBlasHandle;

513:     PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
514:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle));
515:     PetscCall(PetscLogGpuTimeBegin());
516:     PetscCallCUPMBLAS(cupmBlasXaxpy(cupmBlasHandle, n, cupmScalarPtrCast(&alpha), DeviceArrayRead(dctx, xin), 1, DeviceArrayReadWrite(dctx, yin), 1));
517:     PetscCall(PetscLogGpuTimeEnd());
518:     PetscCall(PetscLogGpuFlops(2 * n));
519:     if (n > 0) PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
520:   } else {
521:     PetscCall(VecAXPY_Seq(yin, alpha, xin));
522:   }
523:   PetscFunctionReturn(PETSC_SUCCESS);
524: }

526: // v->ops->axpy
527: template <device::cupm::DeviceType T>
528: inline PetscErrorCode VecSeq_CUPM<T>::AXPY(Vec yin, PetscScalar alpha, Vec xin) noexcept
529: {
530:   PetscFunctionBegin;
531:   PetscCall(AXPYAsync(yin, alpha, xin, nullptr));
532:   PetscFunctionReturn(PETSC_SUCCESS);
533: }

535: namespace detail
536: {

538: struct divides {
539:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &lhs, const PetscScalar &rhs) const noexcept { return rhs == PetscScalar{0.0} ? rhs : lhs / rhs; }
540: };

542: } // namespace detail

544: // VecPointwiseDivideAsync_Private
545: template <device::cupm::DeviceType T>
546: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseDivideAsync(Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
547: {
548:   PetscFunctionBegin;
549:   PetscCall(PointwiseBinaryDispatch_(VecPointwiseDivide_Seq, detail::divides{}, wout, xin, yin, dctx));
550:   PetscFunctionReturn(PETSC_SUCCESS);
551: }

553: // v->ops->pointwisedivide
554: template <device::cupm::DeviceType T>
555: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseDivide(Vec wout, Vec xin, Vec yin) noexcept
556: {
557:   PetscFunctionBegin;
558:   PetscCall(PointwiseDivideAsync(wout, xin, yin, nullptr));
559:   PetscFunctionReturn(PETSC_SUCCESS);
560: }

562: // VecPointwiseMultAsync_Private
563: template <device::cupm::DeviceType T>
564: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMultAsync(Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
565: {
566:   PetscFunctionBegin;
567:   PetscCall(PointwiseBinaryDispatch_(VecPointwiseMult_Seq, thrust::multiplies<PetscScalar>{}, wout, xin, yin, dctx));
568:   PetscFunctionReturn(PETSC_SUCCESS);
569: }

571: // v->ops->pointwisemult
572: template <device::cupm::DeviceType T>
573: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMult(Vec wout, Vec xin, Vec yin) noexcept
574: {
575:   PetscFunctionBegin;
576:   PetscCall(PointwiseMultAsync(wout, xin, yin, nullptr));
577:   PetscFunctionReturn(PETSC_SUCCESS);
578: }

580: namespace detail
581: {

583: struct MaximumRealPart {
584:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &lhs, const PetscScalar &rhs) const noexcept { return thrust::maximum<PetscReal>{}(PetscRealPart(lhs), PetscRealPart(rhs)); }
585: };

587: } // namespace detail

589: // VecPointwiseMaxAsync_Private
590: template <device::cupm::DeviceType T>
591: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMaxAsync(Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
592: {
593:   PetscFunctionBegin;
594:   PetscCall(PointwiseBinaryDispatch_(VecPointwiseMax_Seq, detail::MaximumRealPart{}, wout, xin, yin, dctx));
595:   PetscFunctionReturn(PETSC_SUCCESS);
596: }

598: // v->ops->pointwisemax
599: template <device::cupm::DeviceType T>
600: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMax(Vec wout, Vec xin, Vec yin) noexcept
601: {
602:   PetscFunctionBegin;
603:   PetscCall(PointwiseMaxAsync(wout, xin, yin, nullptr));
604:   PetscFunctionReturn(PETSC_SUCCESS);
605: }

607: namespace detail
608: {

610: struct MaximumAbsoluteValue {
611:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &lhs, const PetscScalar &rhs) const noexcept { return thrust::maximum<PetscReal>{}(PetscAbsScalar(lhs), PetscAbsScalar(rhs)); }
612: };

614: } // namespace detail

616: // VecPointwiseMaxAbsAsync_Private
617: template <device::cupm::DeviceType T>
618: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMaxAbsAsync(Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
619: {
620:   PetscFunctionBegin;
621:   PetscCall(PointwiseBinaryDispatch_(VecPointwiseMaxAbs_Seq, detail::MaximumAbsoluteValue{}, wout, xin, yin, dctx));
622:   PetscFunctionReturn(PETSC_SUCCESS);
623: }

625: // v->ops->pointwisemaxabs
626: template <device::cupm::DeviceType T>
627: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMaxAbs(Vec wout, Vec xin, Vec yin) noexcept
628: {
629:   PetscFunctionBegin;
630:   PetscCall(PointwiseMaxAbsAsync(wout, xin, yin, nullptr));
631:   PetscFunctionReturn(PETSC_SUCCESS);
632: }

634: namespace detail
635: {

637: struct MinimumRealPart {
638:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &lhs, const PetscScalar &rhs) const noexcept { return thrust::minimum<PetscReal>{}(PetscRealPart(lhs), PetscRealPart(rhs)); }
639: };

641: } // namespace detail

643: // VecPointwiseMinAsync_Private
644: template <device::cupm::DeviceType T>
645: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMinAsync(Vec wout, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
646: {
647:   PetscFunctionBegin;
648:   PetscCall(PointwiseBinaryDispatch_(VecPointwiseMin_Seq, detail::MinimumRealPart{}, wout, xin, yin, dctx));
649:   PetscFunctionReturn(PETSC_SUCCESS);
650: }

652: // v->ops->pointwisemin
653: template <device::cupm::DeviceType T>
654: inline PetscErrorCode VecSeq_CUPM<T>::PointwiseMin(Vec wout, Vec xin, Vec yin) noexcept
655: {
656:   PetscFunctionBegin;
657:   PetscCall(PointwiseMinAsync(wout, xin, yin, nullptr));
658:   PetscFunctionReturn(PETSC_SUCCESS);
659: }

661: namespace detail
662: {

664: struct Reciprocal {
665:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &s) const noexcept
666:   {
667:     // yes all of this verbosity is needed because sometimes PetscScalar is a thrust::complex
668:     // and then it matters whether we do s ? true : false vs s == 0, as well as whether we wrap
669:     // everything in PetscScalar...
670:     return s == PetscScalar{0.0} ? s : PetscScalar{1.0} / s;
671:   }
672: };

674: } // namespace detail

676: // VecReciprocalAsync_Private
677: template <device::cupm::DeviceType T>
678: inline PetscErrorCode VecSeq_CUPM<T>::ReciprocalAsync(Vec xin, PetscDeviceContext dctx) noexcept
679: {
680:   PetscFunctionBegin;
681:   PetscCall(PointwiseUnary_(detail::Reciprocal{}, xin, nullptr, dctx));
682:   PetscFunctionReturn(PETSC_SUCCESS);
683: }

685: // v->ops->reciprocal
686: template <device::cupm::DeviceType T>
687: inline PetscErrorCode VecSeq_CUPM<T>::Reciprocal(Vec xin) noexcept
688: {
689:   PetscFunctionBegin;
690:   PetscCall(ReciprocalAsync(xin, nullptr));
691:   PetscFunctionReturn(PETSC_SUCCESS);
692: }

694: namespace detail
695: {

697: struct AbsoluteValue {
698:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &s) const noexcept { return PetscAbsScalar(s); }
699: };

701: } // namespace detail

703: // VecAbsAsync_Private
704: template <device::cupm::DeviceType T>
705: inline PetscErrorCode VecSeq_CUPM<T>::AbsAsync(Vec xin, PetscDeviceContext dctx) noexcept
706: {
707:   PetscFunctionBegin;
708:   PetscCall(PointwiseUnary_(detail::AbsoluteValue{}, xin, nullptr, dctx));
709:   PetscFunctionReturn(PETSC_SUCCESS);
710: }

712: // v->ops->abs
713: template <device::cupm::DeviceType T>
714: inline PetscErrorCode VecSeq_CUPM<T>::Abs(Vec xin) noexcept
715: {
716:   PetscFunctionBegin;
717:   PetscCall(AbsAsync(xin, nullptr));
718:   PetscFunctionReturn(PETSC_SUCCESS);
719: }

721: namespace detail
722: {

724: struct SquareRootAbsoluteValue {
725:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &s) const noexcept { return PetscSqrtReal(PetscAbsScalar(s)); }
726: };

728: } // namespace detail

730: // VecSqrtAbsAsync_Private
731: template <device::cupm::DeviceType T>
732: inline PetscErrorCode VecSeq_CUPM<T>::SqrtAbsAsync(Vec xin, PetscDeviceContext dctx) noexcept
733: {
734:   PetscFunctionBegin;
735:   PetscCall(PointwiseUnary_(detail::SquareRootAbsoluteValue{}, xin, nullptr, dctx));
736:   PetscFunctionReturn(PETSC_SUCCESS);
737: }

739: // v->ops->sqrt
740: template <device::cupm::DeviceType T>
741: inline PetscErrorCode VecSeq_CUPM<T>::SqrtAbs(Vec xin) noexcept
742: {
743:   PetscFunctionBegin;
744:   PetscCall(SqrtAbsAsync(xin, nullptr));
745:   PetscFunctionReturn(PETSC_SUCCESS);
746: }

748: namespace detail
749: {

751: struct Exponent {
752:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &s) const noexcept { return PetscExpScalar(s); }
753: };

755: } // namespace detail

757: // VecExpAsync_Private
758: template <device::cupm::DeviceType T>
759: inline PetscErrorCode VecSeq_CUPM<T>::ExpAsync(Vec xin, PetscDeviceContext dctx) noexcept
760: {
761:   PetscFunctionBegin;
762:   PetscCall(PointwiseUnary_(detail::Exponent{}, xin, nullptr, dctx));
763:   PetscFunctionReturn(PETSC_SUCCESS);
764: }

766: // v->ops->exp
767: template <device::cupm::DeviceType T>
768: inline PetscErrorCode VecSeq_CUPM<T>::Exp(Vec xin) noexcept
769: {
770:   PetscFunctionBegin;
771:   PetscCall(ExpAsync(xin, nullptr));
772:   PetscFunctionReturn(PETSC_SUCCESS);
773: }

775: namespace detail
776: {

778: struct Logarithm {
779:   PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &s) const noexcept { return PetscLogScalar(s); }
780: };

782: } // namespace detail

784: // VecLogAsync_Private
785: template <device::cupm::DeviceType T>
786: inline PetscErrorCode VecSeq_CUPM<T>::LogAsync(Vec xin, PetscDeviceContext dctx) noexcept
787: {
788:   PetscFunctionBegin;
789:   PetscCall(PointwiseUnary_(detail::Logarithm{}, xin, nullptr, dctx));
790:   PetscFunctionReturn(PETSC_SUCCESS);
791: }

793: // v->ops->log
794: template <device::cupm::DeviceType T>
795: inline PetscErrorCode VecSeq_CUPM<T>::Log(Vec xin) noexcept
796: {
797:   PetscFunctionBegin;
798:   PetscCall(LogAsync(xin, nullptr));
799:   PetscFunctionReturn(PETSC_SUCCESS);
800: }

802: // v->ops->waxpy
803: template <device::cupm::DeviceType T>
804: inline PetscErrorCode VecSeq_CUPM<T>::WAXPYAsync(Vec win, PetscScalar alpha, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
805: {
806:   PetscFunctionBegin;
807:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
808:   if (alpha == PetscScalar(0.0)) {
809:     PetscCall(CopyAsync(yin, win, dctx));
810:   } else if (const auto n = static_cast<cupmBlasInt_t>(win->map->n)) {
811:     cupmBlasHandle_t cupmBlasHandle;
812:     cupmStream_t     stream;

814:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle, NULL, &stream));
815:     {
816:       const auto wptr = DeviceArrayWrite(dctx, win);

818:       PetscCall(PetscLogGpuTimeBegin());
819:       PetscCall(PetscCUPMMemcpyAsync(wptr.data(), DeviceArrayRead(dctx, yin).data(), n, cupmMemcpyDeviceToDevice, stream, true));
820:       PetscCallCUPMBLAS(cupmBlasXaxpy(cupmBlasHandle, n, cupmScalarPtrCast(&alpha), DeviceArrayRead(dctx, xin), 1, wptr.cupmdata(), 1));
821:       PetscCall(PetscLogGpuTimeEnd());
822:     }
823:     PetscCall(PetscLogGpuFlops(2 * n));
824:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
825:   }
826:   PetscFunctionReturn(PETSC_SUCCESS);
827: }

829: // v->ops->waxpy
830: template <device::cupm::DeviceType T>
831: inline PetscErrorCode VecSeq_CUPM<T>::WAXPY(Vec win, PetscScalar alpha, Vec xin, Vec yin) noexcept
832: {
833:   PetscFunctionBegin;
834:   PetscCall(WAXPYAsync(win, alpha, xin, yin, nullptr));
835:   PetscFunctionReturn(PETSC_SUCCESS);
836: }

838: namespace kernels
839: {

841: template <typename... Args>
842: PETSC_KERNEL_DECL static void MAXPY_kernel(const PetscInt size, PetscScalar *PETSC_RESTRICT xptr, const PetscScalar *PETSC_RESTRICT aptr, Args... yptr)
843: {
844:   constexpr int      N        = sizeof...(Args);
845:   const auto         tx       = threadIdx.x;
846:   const PetscScalar *yptr_p[] = {yptr...};

848:   PETSC_SHAREDMEM_DECL PetscScalar aptr_shmem[N];

850:   // load a to shared memory
851:   if (tx < N) aptr_shmem[tx] = aptr[tx];
852:   __syncthreads();

854:   ::Petsc::device::cupm::kernels::util::grid_stride_1D(size, [&](PetscInt i) {
855:   // these may look the same but give different results!
856: #if 0
857:     PetscScalar sum = 0.0;

859:   #pragma unroll
860:     for (auto j = 0; j < N; ++j) sum += aptr_shmem[j]*yptr_p[j][i];
861:     xptr[i] += sum;
862: #else
863:     auto sum = xptr[i];

865:   #pragma unroll
866:     for (auto j = 0; j < N; ++j) sum += aptr_shmem[j] * yptr_p[j][i];
867:     xptr[i] = sum;
868: #endif
869:   });
870:   return;
871: }

873: } // namespace kernels

875: namespace detail
876: {

878: // a helper-struct to gobble the size_t input, it is used with template parameter pack
879: // expansion such that
880: // typename repeat_type<MyType, IdxParamPack>...
881: // expands to
882: // MyType, MyType, MyType, ... [repeated sizeof...(IdxParamPack) times]
883: template <typename T, std::size_t>
884: struct repeat_type {
885:   using type = T;
886: };

888: } // namespace detail

890: template <device::cupm::DeviceType T>
891: template <std::size_t... Idx>
892: inline PetscErrorCode VecSeq_CUPM<T>::MAXPY_kernel_dispatch_(PetscDeviceContext dctx, cupmStream_t stream, PetscScalar *xptr, const PetscScalar *aptr, const Vec *yin, PetscInt size, util::index_sequence<Idx...>) noexcept
893: {
894:   PetscFunctionBegin;
895:   // clang-format off
896:   PetscCall(
897:     PetscCUPMLaunchKernel1D(
898:       size, 0, stream,
899:       kernels::MAXPY_kernel<typename detail::repeat_type<const PetscScalar *, Idx>::type...>,
900:       size, xptr, aptr, DeviceArrayRead(dctx, yin[Idx]).data()...
901:     )
902:   );
903:   // clang-format on
904:   PetscFunctionReturn(PETSC_SUCCESS);
905: }

907: template <device::cupm::DeviceType T>
908: template <int N>
909: inline PetscErrorCode VecSeq_CUPM<T>::MAXPY_kernel_dispatch_(PetscDeviceContext dctx, cupmStream_t stream, PetscScalar *xptr, const PetscScalar *aptr, const Vec *yin, PetscInt size, PetscInt &yidx) noexcept
910: {
911:   PetscFunctionBegin;
912:   PetscCall(MAXPY_kernel_dispatch_(dctx, stream, xptr, aptr + yidx, yin + yidx, size, util::make_index_sequence<N>{}));
913:   yidx += N;
914:   PetscFunctionReturn(PETSC_SUCCESS);
915: }

917: // VecMAXPYAsync_Private
918: template <device::cupm::DeviceType T>
919: inline PetscErrorCode VecSeq_CUPM<T>::MAXPYAsync(Vec xin, PetscInt nv, const PetscScalar *alpha, Vec *yin, PetscDeviceContext dctx) noexcept
920: {
921:   const auto   n = xin->map->n;
922:   cupmStream_t stream;

924:   PetscFunctionBegin;
925:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
926:   PetscCall(GetHandlesFrom_(dctx, &stream));
927:   {
928:     const auto   xptr    = DeviceArrayReadWrite(dctx, xin);
929:     PetscScalar *d_alpha = nullptr;
930:     PetscInt     yidx    = 0;

932:     // placement of early-return is deliberate, we would like to capture the
933:     // DeviceArrayReadWrite() call (which calls PetscObjectStateIncreate()) before we bail
934:     if (!n || !nv) PetscFunctionReturn(PETSC_SUCCESS);
935:     PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_CUPM(), nv, &d_alpha));
936:     PetscCall(PetscCUPMMemcpyAsync(d_alpha, alpha, nv, cupmMemcpyHostToDevice, stream));
937:     PetscCall(PetscLogGpuTimeBegin());
938:     do {
939:       switch (nv - yidx) {
940:       case 7:
941:         PetscCall(MAXPY_kernel_dispatch_<7>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
942:         break;
943:       case 6:
944:         PetscCall(MAXPY_kernel_dispatch_<6>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
945:         break;
946:       case 5:
947:         PetscCall(MAXPY_kernel_dispatch_<5>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
948:         break;
949:       case 4:
950:         PetscCall(MAXPY_kernel_dispatch_<4>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
951:         break;
952:       case 3:
953:         PetscCall(MAXPY_kernel_dispatch_<3>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
954:         break;
955:       case 2:
956:         PetscCall(MAXPY_kernel_dispatch_<2>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
957:         break;
958:       case 1:
959:         PetscCall(MAXPY_kernel_dispatch_<1>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
960:         break;
961:       default: // 8 or more
962:         PetscCall(MAXPY_kernel_dispatch_<8>(dctx, stream, xptr.data(), d_alpha, yin, n, yidx));
963:         break;
964:       }
965:     } while (yidx < nv);
966:     PetscCall(PetscLogGpuTimeEnd());
967:     PetscCall(PetscDeviceFree(dctx, d_alpha));
968:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
969:   }
970:   PetscCall(PetscLogGpuFlops(nv * 2 * n));
971:   PetscFunctionReturn(PETSC_SUCCESS);
972: }

974: // v->ops->maxpy
975: template <device::cupm::DeviceType T>
976: inline PetscErrorCode VecSeq_CUPM<T>::MAXPY(Vec xin, PetscInt nv, const PetscScalar *alpha, Vec *yin) noexcept
977: {
978:   PetscFunctionBegin;
979:   PetscCall(MAXPYAsync(xin, nv, alpha, yin, nullptr));
980:   PetscFunctionReturn(PETSC_SUCCESS);
981: }

983: template <device::cupm::DeviceType T>
984: inline PetscErrorCode VecSeq_CUPM<T>::Dot(Vec xin, Vec yin, PetscScalar *z) noexcept
985: {
986:   PetscFunctionBegin;
987:   if (const auto n = static_cast<cupmBlasInt_t>(xin->map->n)) {
988:     PetscDeviceContext dctx;
989:     cupmBlasHandle_t   cupmBlasHandle;

991:     PetscCall(GetHandles_(&dctx, &cupmBlasHandle));
992:     // arguments y, x are reversed because BLAS complex conjugates the first argument, PETSc the
993:     // second
994:     PetscCall(PetscLogGpuTimeBegin());
995:     PetscCallCUPMBLAS(cupmBlasXdot(cupmBlasHandle, n, DeviceArrayRead(dctx, yin), 1, DeviceArrayRead(dctx, xin), 1, cupmScalarPtrCast(z)));
996:     PetscCall(PetscLogGpuTimeEnd());
997:     PetscCall(PetscLogGpuFlops(2 * n - 1));
998:   } else {
999:     *z = 0.0;
1000:   }
1001:   PetscFunctionReturn(PETSC_SUCCESS);
1002: }

1004: #define MDOT_WORKGROUP_NUM  128
1005: #define MDOT_WORKGROUP_SIZE MDOT_WORKGROUP_NUM

1007: namespace kernels
1008: {

1010: PETSC_DEVICE_INLINE_DECL static PetscInt EntriesPerGroup(const PetscInt size) noexcept
1011: {
1012:   const auto group_entries = (size - 1) / gridDim.x + 1;
1013:   // for very small vectors, a group should still do some work
1014:   return group_entries ? group_entries : 1;
1015: }

1017: template <typename... ConstPetscScalarPointer>
1018: PETSC_KERNEL_DECL static void MDot_kernel(const PetscScalar *PETSC_RESTRICT x, const PetscInt size, PetscScalar *PETSC_RESTRICT results, ConstPetscScalarPointer... y)
1019: {
1020:   constexpr int      N        = sizeof...(ConstPetscScalarPointer);
1021:   const PetscScalar *ylocal[] = {y...};
1022:   PetscScalar        sumlocal[N];

1024:   PETSC_SHAREDMEM_DECL PetscScalar shmem[N * MDOT_WORKGROUP_SIZE];

1026:   // HIP -- for whatever reason -- has threadIdx, blockIdx, blockDim, and gridDim as separate
1027:   // types, so each of these go on separate lines...
1028:   const auto tx       = threadIdx.x;
1029:   const auto bx       = blockIdx.x;
1030:   const auto bdx      = blockDim.x;
1031:   const auto gdx      = gridDim.x;
1032:   const auto worksize = EntriesPerGroup(size);
1033:   const auto begin    = tx + bx * worksize;
1034:   const auto end      = min((bx + 1) * worksize, size);

1036: #pragma unroll
1037:   for (auto i = 0; i < N; ++i) sumlocal[i] = 0;

1039:   for (auto i = begin; i < end; i += bdx) {
1040:     const auto xi = x[i]; // load only once from global memory!

1042: #pragma unroll
1043:     for (auto j = 0; j < N; ++j) sumlocal[j] += ylocal[j][i] * xi;
1044:   }

1046: #pragma unroll
1047:   for (auto i = 0; i < N; ++i) shmem[tx + i * MDOT_WORKGROUP_SIZE] = sumlocal[i];

1049:   // parallel reduction
1050:   for (auto stride = bdx / 2; stride > 0; stride /= 2) {
1051:     __syncthreads();
1052:     if (tx < stride) {
1053: #pragma unroll
1054:       for (auto i = 0; i < N; ++i) shmem[tx + i * MDOT_WORKGROUP_SIZE] += shmem[tx + stride + i * MDOT_WORKGROUP_SIZE];
1055:     }
1056:   }
1057:   // bottom N threads per block write to global memory
1058:   // REVIEW ME: I am ~pretty~ sure we don't need another __syncthreads() here since each thread
1059:   // writes to the same sections in the above loop that it is about to read from below, but
1060:   // running this under the racecheck tool of cuda-memcheck reports a write-after-write hazard.
1061:   __syncthreads();
1062:   if (tx < N) results[bx + tx * gdx] = shmem[tx * MDOT_WORKGROUP_SIZE];
1063:   return;
1064: }

1066: namespace
1067: {

1069: PETSC_KERNEL_DECL void sum_kernel(const PetscInt size, PetscScalar *PETSC_RESTRICT results)
1070: {
1071:   int         local_i = 0;
1072:   PetscScalar local_results[8];

1074:   // each thread sums up MDOT_WORKGROUP_NUM entries of the result, storing it in a local buffer
1075:   //
1076:   // *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*
1077:   // | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | ...
1078:   // *-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*-*
1079:   //  |  ______________________________________________________/
1080:   //  | /            <- MDOT_WORKGROUP_NUM ->
1081:   //  |/
1082:   //  +
1083:   //  v
1084:   // *-*-*
1085:   // | | | ...
1086:   // *-*-*
1087:   //
1088:   ::Petsc::device::cupm::kernels::util::grid_stride_1D(size, [&](PetscInt i) {
1089:     PetscScalar z_sum = 0;

1091:     for (auto j = i * MDOT_WORKGROUP_SIZE; j < (i + 1) * MDOT_WORKGROUP_SIZE; ++j) z_sum += results[j];
1092:     local_results[local_i++] = z_sum;
1093:   });
1094:   // if we needed more than 1 workgroup to handle the vector we should sync since other threads
1095:   // may currently be reading from results
1096:   if (size >= MDOT_WORKGROUP_SIZE) __syncthreads();
1097:   // Local buffer is now written to global memory
1098:   ::Petsc::device::cupm::kernels::util::grid_stride_1D(size, [&](PetscInt i) {
1099:     const auto j = --local_i;

1101:     if (j >= 0) results[i] = local_results[j];
1102:   });
1103:   return;
1104: }

1106: } // namespace

1108: #if PetscDefined(USING_HCC)
1109: namespace do_not_use
1110: {

1112: inline void silence_warning_function_sum_kernel_is_not_needed_and_will_not_be_emitted()
1113: {
1114:   (void)sum_kernel;
1115: }

1117: } // namespace do_not_use
1118: #endif

1120: } // namespace kernels

1122: template <device::cupm::DeviceType T>
1123: template <std::size_t... Idx>
1124: inline PetscErrorCode VecSeq_CUPM<T>::MDot_kernel_dispatch_(PetscDeviceContext dctx, cupmStream_t stream, const PetscScalar *xarr, const Vec yin[], PetscInt size, PetscScalar *results, util::index_sequence<Idx...>) noexcept
1125: {
1126:   PetscFunctionBegin;
1127:   // REVIEW ME: convert this kernel launch to PetscCUPMLaunchKernel1D(), it currently launches
1128:   // 128 blocks of 128 threads every time which may be wasteful
1129:   // clang-format off
1130:   PetscCallCUPM(
1131:     cupmLaunchKernel(
1132:       kernels::MDot_kernel<typename detail::repeat_type<const PetscScalar *, Idx>::type...>,
1133:       MDOT_WORKGROUP_NUM, MDOT_WORKGROUP_SIZE, 0, stream,
1134:       xarr, size, results, DeviceArrayRead(dctx, yin[Idx]).data()...
1135:     )
1136:   );
1137:   // clang-format on
1138:   PetscFunctionReturn(PETSC_SUCCESS);
1139: }

1141: template <device::cupm::DeviceType T>
1142: template <int N>
1143: inline PetscErrorCode VecSeq_CUPM<T>::MDot_kernel_dispatch_(PetscDeviceContext dctx, cupmStream_t stream, const PetscScalar *xarr, const Vec yin[], PetscInt size, PetscScalar *results, PetscInt &yidx) noexcept
1144: {
1145:   PetscFunctionBegin;
1146:   PetscCall(MDot_kernel_dispatch_(dctx, stream, xarr, yin + yidx, size, results + yidx * MDOT_WORKGROUP_NUM, util::make_index_sequence<N>{}));
1147:   yidx += N;
1148:   PetscFunctionReturn(PETSC_SUCCESS);
1149: }

1151: template <device::cupm::DeviceType T>
1152: inline PetscErrorCode VecSeq_CUPM<T>::MDot_(std::false_type, Vec xin, PetscInt nv, const Vec yin[], PetscScalar *z, PetscDeviceContext dctx) noexcept
1153: {
1154:   // the largest possible size of a batch
1155:   constexpr PetscInt batchsize = 8;
1156:   // how many sub streams to create, if nv <= batchsize we can do this without looping, so we
1157:   // do not create substreams. Note we don't create more than 8 streams, in practice we could
1158:   // not get more parallelism with higher numbers.
1159:   const auto num_sub_streams = nv > batchsize ? std::min((nv + batchsize) / batchsize, batchsize) : 0;
1160:   const auto n               = xin->map->n;
1161:   // number of vectors that we handle via the batches. note any singletons are handled by
1162:   // cublas, hence the nv-1.
1163:   const auto   nvbatch = ((nv % batchsize) == 1) ? nv - 1 : nv;
1164:   const auto   nwork   = nvbatch * MDOT_WORKGROUP_NUM;
1165:   PetscScalar *d_results;
1166:   cupmStream_t stream;

1168:   PetscFunctionBegin;
1169:   PetscCall(GetHandlesFrom_(dctx, &stream));
1170:   // allocate scratchpad memory for the results of individual work groups
1171:   PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_CUPM(), nwork, &d_results));
1172:   {
1173:     const auto          xptr       = DeviceArrayRead(dctx, xin);
1174:     PetscInt            yidx       = 0;
1175:     auto                subidx     = 0;
1176:     auto                cur_stream = stream;
1177:     auto                cur_ctx    = dctx;
1178:     PetscDeviceContext *sub        = nullptr;
1179:     PetscStreamType     stype;

1181:     // REVIEW ME: maybe PetscDeviceContextFork() should insert dctx into the first entry of
1182:     // sub. Ideally the parent context should also join in on the fork, but it is extremely
1183:     // fiddly to do so presently
1184:     PetscCall(PetscDeviceContextGetStreamType(dctx, &stype));
1185:     if (stype == PETSC_STREAM_GLOBAL_BLOCKING) stype = PETSC_STREAM_DEFAULT_BLOCKING;
1186:     // If we have a globally blocking stream create nonblocking streams instead (as we can
1187:     // locally exploit the parallelism). Otherwise use the prescribed stream type.
1188:     PetscCall(PetscDeviceContextForkWithStreamType(dctx, stype, num_sub_streams, &sub));
1189:     PetscCall(PetscLogGpuTimeBegin());
1190:     do {
1191:       if (num_sub_streams) {
1192:         cur_ctx = sub[subidx++ % num_sub_streams];
1193:         PetscCall(GetHandlesFrom_(cur_ctx, &cur_stream));
1194:       }
1195:       // REVIEW ME: Should probably try and load-balance these. Consider the case where nv = 9;
1196:       // it is very likely better to do 4+5 rather than 8+1
1197:       switch (nv - yidx) {
1198:       case 7:
1199:         PetscCall(MDot_kernel_dispatch_<7>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1200:         break;
1201:       case 6:
1202:         PetscCall(MDot_kernel_dispatch_<6>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1203:         break;
1204:       case 5:
1205:         PetscCall(MDot_kernel_dispatch_<5>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1206:         break;
1207:       case 4:
1208:         PetscCall(MDot_kernel_dispatch_<4>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1209:         break;
1210:       case 3:
1211:         PetscCall(MDot_kernel_dispatch_<3>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1212:         break;
1213:       case 2:
1214:         PetscCall(MDot_kernel_dispatch_<2>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1215:         break;
1216:       case 1: {
1217:         cupmBlasHandle_t cupmBlasHandle;

1219:         PetscCall(GetHandlesFrom_(cur_ctx, &cupmBlasHandle));
1220:         PetscCallCUPMBLAS(cupmBlasXdot(cupmBlasHandle, static_cast<cupmBlasInt_t>(n), DeviceArrayRead(cur_ctx, yin[yidx]).cupmdata(), 1, xptr.cupmdata(), 1, cupmScalarPtrCast(z + yidx)));
1221:         ++yidx;
1222:       } break;
1223:       default: // 8 or more
1224:         PetscCall(MDot_kernel_dispatch_<8>(cur_ctx, cur_stream, xptr.data(), yin, n, d_results, yidx));
1225:         break;
1226:       }
1227:     } while (yidx < nv);
1228:     PetscCall(PetscLogGpuTimeEnd());
1229:     PetscCall(PetscDeviceContextJoin(dctx, num_sub_streams, PETSC_DEVICE_CONTEXT_JOIN_DESTROY, &sub));
1230:   }

1232:   PetscCall(PetscCUPMLaunchKernel1D(nvbatch, 0, stream, kernels::sum_kernel, nvbatch, d_results));
1233:   // copy result of device reduction to host
1234:   PetscCall(PetscCUPMMemcpyAsync(z, d_results, nvbatch, cupmMemcpyDeviceToHost, stream));
1235:   // do these now while final reduction is in flight
1236:   PetscCall(PetscLogGpuFlops(nwork));
1237:   PetscCall(PetscDeviceFree(dctx, d_results));
1238:   PetscFunctionReturn(PETSC_SUCCESS);
1239: }

1241: #undef MDOT_WORKGROUP_NUM
1242: #undef MDOT_WORKGROUP_SIZE

1244: template <device::cupm::DeviceType T>
1245: inline PetscErrorCode VecSeq_CUPM<T>::MDot_(std::true_type, Vec xin, PetscInt nv, const Vec yin[], PetscScalar *z, PetscDeviceContext dctx) noexcept
1246: {
1247:   // probably not worth it to run more than 8 of these at a time?
1248:   const auto          n_sub = PetscMin(nv, 8);
1249:   const auto          n     = static_cast<cupmBlasInt_t>(xin->map->n);
1250:   const auto          xptr  = DeviceArrayRead(dctx, xin);
1251:   PetscScalar        *d_z;
1252:   PetscDeviceContext *subctx;
1253:   cupmStream_t        stream;

1255:   PetscFunctionBegin;
1256:   PetscCall(GetHandlesFrom_(dctx, &stream));
1257:   PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_CUPM(), nv, &d_z));
1258:   PetscCall(PetscDeviceContextFork(dctx, n_sub, &subctx));
1259:   PetscCall(PetscLogGpuTimeBegin());
1260:   for (PetscInt i = 0; i < nv; ++i) {
1261:     const auto            sub = subctx[i % n_sub];
1262:     cupmBlasHandle_t      handle;
1263:     cupmBlasPointerMode_t old_mode;

1265:     PetscCall(GetHandlesFrom_(sub, &handle));
1266:     PetscCallCUPMBLAS(cupmBlasGetPointerMode(handle, &old_mode));
1267:     if (old_mode != CUPMBLAS_POINTER_MODE_DEVICE) PetscCallCUPMBLAS(cupmBlasSetPointerMode(handle, CUPMBLAS_POINTER_MODE_DEVICE));
1268:     PetscCallCUPMBLAS(cupmBlasXdot(handle, n, DeviceArrayRead(sub, yin[i]), 1, xptr.cupmdata(), 1, cupmScalarPtrCast(d_z + i)));
1269:     if (old_mode != CUPMBLAS_POINTER_MODE_DEVICE) PetscCallCUPMBLAS(cupmBlasSetPointerMode(handle, old_mode));
1270:   }
1271:   PetscCall(PetscLogGpuTimeEnd());
1272:   PetscCall(PetscDeviceContextJoin(dctx, n_sub, PETSC_DEVICE_CONTEXT_JOIN_DESTROY, &subctx));
1273:   PetscCall(PetscCUPMMemcpyAsync(z, d_z, nv, cupmMemcpyDeviceToHost, stream));
1274:   PetscCall(PetscDeviceFree(dctx, d_z));
1275:   // REVIEW ME: flops?????
1276:   PetscFunctionReturn(PETSC_SUCCESS);
1277: }

1279: // v->ops->mdot
1280: template <device::cupm::DeviceType T>
1281: inline PetscErrorCode VecSeq_CUPM<T>::MDot(Vec xin, PetscInt nv, const Vec yin[], PetscScalar *z) noexcept
1282: {
1283:   PetscFunctionBegin;
1284:   if (PetscUnlikely(nv == 1)) {
1285:     // dot handles nv = 0 correctly
1286:     PetscCall(Dot(xin, const_cast<Vec>(yin[0]), z));
1287:   } else if (const auto n = xin->map->n) {
1288:     PetscDeviceContext dctx;

1290:     PetscCheck(nv > 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Number of vectors provided to %s %" PetscInt_FMT " not positive", PETSC_FUNCTION_NAME, nv);
1291:     PetscCall(GetHandles_(&dctx));
1292:     PetscCall(MDot_(std::integral_constant<bool, PetscDefined(USE_COMPLEX)>{}, xin, nv, yin, z, dctx));
1293:     // REVIEW ME: double count of flops??
1294:     PetscCall(PetscLogGpuFlops(nv * (2 * n - 1)));
1295:     PetscCall(PetscDeviceContextSynchronize(dctx));
1296:   } else {
1297:     PetscCall(PetscArrayzero(z, nv));
1298:   }
1299:   PetscFunctionReturn(PETSC_SUCCESS);
1300: }

1302: // VecSetAsync_Private
1303: template <device::cupm::DeviceType T>
1304: inline PetscErrorCode VecSeq_CUPM<T>::SetAsync(Vec xin, PetscScalar alpha, PetscDeviceContext dctx) noexcept
1305: {
1306:   const auto   n = xin->map->n;
1307:   cupmStream_t stream;

1309:   PetscFunctionBegin;
1310:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1311:   PetscCall(GetHandlesFrom_(dctx, &stream));
1312:   {
1313:     const auto xptr = DeviceArrayWrite(dctx, xin);

1315:     if (alpha == PetscScalar(0.0)) {
1316:       PetscCall(PetscCUPMMemsetAsync(xptr.data(), 0, n, stream));
1317:     } else {
1318:       const auto dptr = thrust::device_pointer_cast(xptr.data());

1320:       PetscCallThrust(THRUST_CALL(thrust::fill, stream, dptr, dptr + n, alpha));
1321:     }
1322:   }
1323:   if (n > 0) PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
1324:   PetscFunctionReturn(PETSC_SUCCESS);
1325: }

1327: // v->ops->set
1328: template <device::cupm::DeviceType T>
1329: inline PetscErrorCode VecSeq_CUPM<T>::Set(Vec xin, PetscScalar alpha) noexcept
1330: {
1331:   PetscFunctionBegin;
1332:   PetscCall(SetAsync(xin, alpha, nullptr));
1333:   PetscFunctionReturn(PETSC_SUCCESS);
1334: }

1336: // VecScaleAsync_Private
1337: template <device::cupm::DeviceType T>
1338: inline PetscErrorCode VecSeq_CUPM<T>::ScaleAsync(Vec xin, PetscScalar alpha, PetscDeviceContext dctx) noexcept
1339: {
1340:   PetscFunctionBegin;
1341:   if (PetscUnlikely(alpha == PetscScalar(1.0))) PetscFunctionReturn(PETSC_SUCCESS);
1342:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1343:   if (PetscUnlikely(alpha == PetscScalar(0.0))) {
1344:     PetscCall(SetAsync(xin, alpha, dctx));
1345:   } else if (const auto n = static_cast<cupmBlasInt_t>(xin->map->n)) {
1346:     cupmBlasHandle_t cupmBlasHandle;

1348:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle));
1349:     PetscCall(PetscLogGpuTimeBegin());
1350:     PetscCallCUPMBLAS(cupmBlasXscal(cupmBlasHandle, n, cupmScalarPtrCast(&alpha), DeviceArrayReadWrite(dctx, xin), 1));
1351:     PetscCall(PetscLogGpuTimeEnd());
1352:     PetscCall(PetscLogGpuFlops(n));
1353:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
1354:   } else {
1355:     PetscCall(MaybeIncrementEmptyLocalVec(xin));
1356:   }
1357:   PetscFunctionReturn(PETSC_SUCCESS);
1358: }

1360: // v->ops->scale
1361: template <device::cupm::DeviceType T>
1362: inline PetscErrorCode VecSeq_CUPM<T>::Scale(Vec xin, PetscScalar alpha) noexcept
1363: {
1364:   PetscFunctionBegin;
1365:   PetscCall(ScaleAsync(xin, alpha, nullptr));
1366:   PetscFunctionReturn(PETSC_SUCCESS);
1367: }

1369: // v->ops->tdot
1370: template <device::cupm::DeviceType T>
1371: inline PetscErrorCode VecSeq_CUPM<T>::TDot(Vec xin, Vec yin, PetscScalar *z) noexcept
1372: {
1373:   PetscFunctionBegin;
1374:   if (const auto n = static_cast<cupmBlasInt_t>(xin->map->n)) {
1375:     PetscDeviceContext dctx;
1376:     cupmBlasHandle_t   cupmBlasHandle;

1378:     PetscCall(GetHandles_(&dctx, &cupmBlasHandle));
1379:     PetscCall(PetscLogGpuTimeBegin());
1380:     PetscCallCUPMBLAS(cupmBlasXdotu(cupmBlasHandle, n, DeviceArrayRead(dctx, xin), 1, DeviceArrayRead(dctx, yin), 1, cupmScalarPtrCast(z)));
1381:     PetscCall(PetscLogGpuTimeEnd());
1382:     PetscCall(PetscLogGpuFlops(2 * n - 1));
1383:   } else {
1384:     *z = 0.0;
1385:   }
1386:   PetscFunctionReturn(PETSC_SUCCESS);
1387: }

1389: // VecCopyAsync_Private
1390: template <device::cupm::DeviceType T>
1391: inline PetscErrorCode VecSeq_CUPM<T>::CopyAsync(Vec xin, Vec yout, PetscDeviceContext dctx) noexcept
1392: {
1393:   PetscFunctionBegin;
1394:   if (xin == yout) PetscFunctionReturn(PETSC_SUCCESS);
1395:   if (const auto n = xin->map->n) {
1396:     const auto xmask = xin->offloadmask;
1397:     // silence buggy gcc warning: mode may be used uninitialized in this function
1398:     auto         mode = cupmMemcpyDeviceToDevice;
1399:     cupmStream_t stream;

1401:     // translate from PetscOffloadMask to cupmMemcpyKind
1402:     PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1403:     switch (const auto ymask = yout->offloadmask) {
1404:     case PETSC_OFFLOAD_UNALLOCATED: {
1405:       PetscBool yiscupm;

1407:       PetscCall(PetscObjectTypeCompareAny(PetscObjectCast(yout), &yiscupm, VECSEQCUPM(), VECMPICUPM(), ""));
1408:       if (yiscupm) {
1409:         mode = PetscOffloadDevice(xmask) ? cupmMemcpyDeviceToDevice : cupmMemcpyHostToHost;
1410:         break;
1411:       }
1412:     } // fall-through if unallocated and not cupm
1413: #if PETSC_CPP_VERSION >= 17
1414:       [[fallthrough]];
1415: #endif
1416:     case PETSC_OFFLOAD_CPU:
1417:       mode = PetscOffloadHost(xmask) ? cupmMemcpyHostToHost : cupmMemcpyDeviceToHost;
1418:       break;
1419:     case PETSC_OFFLOAD_BOTH:
1420:     case PETSC_OFFLOAD_GPU:
1421:       mode = PetscOffloadDevice(xmask) ? cupmMemcpyDeviceToDevice : cupmMemcpyHostToDevice;
1422:       break;
1423:     default:
1424:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Incompatible offload mask %s", PetscOffloadMaskToString(ymask));
1425:     }

1427:     PetscCall(GetHandlesFrom_(dctx, &stream));
1428:     switch (mode) {
1429:     case cupmMemcpyDeviceToDevice: // the best case
1430:     case cupmMemcpyHostToDevice: { // not terrible
1431:       const auto yptr = DeviceArrayWrite(dctx, yout);
1432:       const auto xptr = mode == cupmMemcpyDeviceToDevice ? DeviceArrayRead(dctx, xin).data() : HostArrayRead(dctx, xin).data();

1434:       PetscCall(PetscLogGpuTimeBegin());
1435:       PetscCall(PetscCUPMMemcpyAsync(yptr.data(), xptr, n, mode, stream));
1436:       PetscCall(PetscLogGpuTimeEnd());
1437:     } break;
1438:     case cupmMemcpyDeviceToHost: // not great
1439:     case cupmMemcpyHostToHost: { // worst case
1440:       const auto   xptr = mode == cupmMemcpyDeviceToHost ? DeviceArrayRead(dctx, xin).data() : HostArrayRead(dctx, xin).data();
1441:       PetscScalar *yptr;

1443:       PetscCall(VecGetArrayWrite(yout, &yptr));
1444:       if (mode == cupmMemcpyDeviceToHost) PetscCall(PetscLogGpuTimeBegin());
1445:       PetscCall(PetscCUPMMemcpyAsync(yptr, xptr, n, mode, stream, /* force async */ true));
1446:       if (mode == cupmMemcpyDeviceToHost) PetscCall(PetscLogGpuTimeEnd());
1447:       PetscCall(VecRestoreArrayWrite(yout, &yptr));
1448:     } break;
1449:     default:
1450:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "Unknown cupmMemcpyKind %d", static_cast<int>(mode));
1451:     }
1452:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
1453:   } else {
1454:     PetscCall(MaybeIncrementEmptyLocalVec(yout));
1455:   }
1456:   PetscFunctionReturn(PETSC_SUCCESS);
1457: }

1459: // v->ops->copy
1460: template <device::cupm::DeviceType T>
1461: inline PetscErrorCode VecSeq_CUPM<T>::Copy(Vec xin, Vec yout) noexcept
1462: {
1463:   PetscFunctionBegin;
1464:   PetscCall(CopyAsync(xin, yout, nullptr));
1465:   PetscFunctionReturn(PETSC_SUCCESS);
1466: }

1468: // VecSwapAsync_Private
1469: template <device::cupm::DeviceType T>
1470: inline PetscErrorCode VecSeq_CUPM<T>::SwapAsync(Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
1471: {
1472:   PetscFunctionBegin;
1473:   if (xin == yin) PetscFunctionReturn(PETSC_SUCCESS);
1474:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1475:   if (const auto n = static_cast<cupmBlasInt_t>(xin->map->n)) {
1476:     cupmBlasHandle_t cupmBlasHandle;

1478:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle));
1479:     PetscCall(PetscLogGpuTimeBegin());
1480:     PetscCallCUPMBLAS(cupmBlasXswap(cupmBlasHandle, n, DeviceArrayReadWrite(dctx, xin), 1, DeviceArrayReadWrite(dctx, yin), 1));
1481:     PetscCall(PetscLogGpuTimeEnd());
1482:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
1483:   } else {
1484:     PetscCall(MaybeIncrementEmptyLocalVec(xin));
1485:     PetscCall(MaybeIncrementEmptyLocalVec(yin));
1486:   }
1487:   PetscFunctionReturn(PETSC_SUCCESS);
1488: }

1490: // v->ops->swap
1491: template <device::cupm::DeviceType T>
1492: inline PetscErrorCode VecSeq_CUPM<T>::Swap(Vec xin, Vec yin) noexcept
1493: {
1494:   PetscFunctionBegin;
1495:   PetscCall(SwapAsync(xin, yin, nullptr));
1496:   PetscFunctionReturn(PETSC_SUCCESS);
1497: }

1499: // VecAXPYBYAsync_Private
1500: template <device::cupm::DeviceType T>
1501: inline PetscErrorCode VecSeq_CUPM<T>::AXPBYAsync(Vec yin, PetscScalar alpha, PetscScalar beta, Vec xin, PetscDeviceContext dctx) noexcept
1502: {
1503:   PetscFunctionBegin;
1504:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1505:   if (alpha == PetscScalar(0.0)) {
1506:     PetscCall(ScaleAsync(yin, beta, dctx));
1507:   } else if (beta == PetscScalar(1.0)) {
1508:     PetscCall(AXPYAsync(yin, alpha, xin, dctx));
1509:   } else if (alpha == PetscScalar(1.0)) {
1510:     PetscCall(AYPXAsync(yin, beta, xin, dctx));
1511:   } else if (const auto n = static_cast<cupmBlasInt_t>(yin->map->n)) {
1512:     const auto       betaIsZero = beta == PetscScalar(0.0);
1513:     const auto       aptr       = cupmScalarPtrCast(&alpha);
1514:     cupmBlasHandle_t cupmBlasHandle;

1516:     PetscCall(GetHandlesFrom_(dctx, &cupmBlasHandle));
1517:     {
1518:       const auto xptr = DeviceArrayRead(dctx, xin);

1520:       if (betaIsZero /* beta = 0 */) {
1521:         // here we can get away with purely write-only as we memcpy into it first
1522:         const auto   yptr = DeviceArrayWrite(dctx, yin);
1523:         cupmStream_t stream;

1525:         PetscCall(GetHandlesFrom_(dctx, &stream));
1526:         PetscCall(PetscLogGpuTimeBegin());
1527:         PetscCall(PetscCUPMMemcpyAsync(yptr.data(), xptr.data(), n, cupmMemcpyDeviceToDevice, stream));
1528:         PetscCallCUPMBLAS(cupmBlasXscal(cupmBlasHandle, n, aptr, yptr.cupmdata(), 1));
1529:       } else {
1530:         const auto yptr = DeviceArrayReadWrite(dctx, yin);

1532:         PetscCall(PetscLogGpuTimeBegin());
1533:         PetscCallCUPMBLAS(cupmBlasXscal(cupmBlasHandle, n, cupmScalarPtrCast(&beta), yptr.cupmdata(), 1));
1534:         PetscCallCUPMBLAS(cupmBlasXaxpy(cupmBlasHandle, n, aptr, xptr.cupmdata(), 1, yptr.cupmdata(), 1));
1535:       }
1536:     }
1537:     PetscCall(PetscLogGpuTimeEnd());
1538:     PetscCall(PetscLogGpuFlops((betaIsZero ? 1 : 3) * n));
1539:     PetscCall(PetscDeviceContextSynchronizeIfGlobalBlocking_Internal(dctx));
1540:   } else {
1541:     PetscCall(MaybeIncrementEmptyLocalVec(yin));
1542:   }
1543:   PetscFunctionReturn(PETSC_SUCCESS);
1544: }

1546: // v->ops->axpby
1547: template <device::cupm::DeviceType T>
1548: inline PetscErrorCode VecSeq_CUPM<T>::AXPBY(Vec yin, PetscScalar alpha, PetscScalar beta, Vec xin) noexcept
1549: {
1550:   PetscFunctionBegin;
1551:   PetscCall(AXPBYAsync(yin, alpha, beta, xin, nullptr));
1552:   PetscFunctionReturn(PETSC_SUCCESS);
1553: }

1555: // VecAXPBYPCZAsync_Private
1556: template <device::cupm::DeviceType T>
1557: inline PetscErrorCode VecSeq_CUPM<T>::AXPBYPCZAsync(Vec zin, PetscScalar alpha, PetscScalar beta, PetscScalar gamma, Vec xin, Vec yin, PetscDeviceContext dctx) noexcept
1558: {
1559:   PetscFunctionBegin;
1560:   PetscCall(PetscDeviceContextGetOptionalNullContext_Internal(&dctx));
1561:   if (gamma != PetscScalar(1.0)) PetscCall(ScaleAsync(zin, gamma, dctx));
1562:   PetscCall(AXPYAsync(zin, alpha, xin, dctx));
1563:   PetscCall(AXPYAsync(zin, beta, yin, dctx));
1564:   PetscFunctionReturn(PETSC_SUCCESS);
1565: }

1567: // v->ops->axpbypcz
1568: template <device::cupm::DeviceType T>
1569: inline PetscErrorCode VecSeq_CUPM<T>::AXPBYPCZ(Vec zin, PetscScalar alpha, PetscScalar beta, PetscScalar gamma, Vec xin, Vec yin) noexcept
1570: {
1571:   PetscFunctionBegin;
1572:   PetscCall(AXPBYPCZAsync(zin, alpha, beta, gamma, xin, yin, nullptr));
1573:   PetscFunctionReturn(PETSC_SUCCESS);
1574: }

1576: // v->ops->norm
1577: template <device::cupm::DeviceType T>
1578: inline PetscErrorCode VecSeq_CUPM<T>::Norm(Vec xin, NormType type, PetscReal *z) noexcept
1579: {
1580:   PetscDeviceContext dctx;
1581:   cupmBlasHandle_t   cupmBlasHandle;

1583:   PetscFunctionBegin;
1584:   PetscCall(GetHandles_(&dctx, &cupmBlasHandle));
1585:   if (const auto n = static_cast<cupmBlasInt_t>(xin->map->n)) {
1586:     const auto xptr      = DeviceArrayRead(dctx, xin);
1587:     PetscInt   flopCount = 0;

1589:     PetscCall(PetscLogGpuTimeBegin());
1590:     switch (type) {
1591:     case NORM_1_AND_2:
1592:     case NORM_1:
1593:       PetscCallCUPMBLAS(cupmBlasXasum(cupmBlasHandle, n, xptr.cupmdata(), 1, cupmRealPtrCast(z)));
1594:       flopCount = std::max(n - 1, 0);
1595:       if (type == NORM_1) break;
1596:       ++z; // fall-through
1597: #if PETSC_CPP_VERSION >= 17
1598:       [[fallthrough]];
1599: #endif
1600:     case NORM_2:
1601:     case NORM_FROBENIUS:
1602:       PetscCallCUPMBLAS(cupmBlasXnrm2(cupmBlasHandle, n, xptr.cupmdata(), 1, cupmRealPtrCast(z)));
1603:       flopCount += std::max(2 * n - 1, 0); // += in case we've fallen through from NORM_1_AND_2
1604:       break;
1605:     case NORM_INFINITY: {
1606:       cupmBlasInt_t max_loc = 0;
1607:       PetscScalar   xv      = 0.;
1608:       cupmStream_t  stream;

1610:       PetscCall(GetHandlesFrom_(dctx, &stream));
1611:       PetscCallCUPMBLAS(cupmBlasXamax(cupmBlasHandle, n, xptr.cupmdata(), 1, &max_loc));
1612:       PetscCall(PetscCUPMMemcpyAsync(&xv, xptr.data() + max_loc - 1, 1, cupmMemcpyDeviceToHost, stream));
1613:       *z = PetscAbsScalar(xv);
1614:       // REVIEW ME: flopCount = ???
1615:     } break;
1616:     }
1617:     PetscCall(PetscLogGpuTimeEnd());
1618:     PetscCall(PetscLogGpuFlops(flopCount));
1619:   } else {
1620:     z[0]                    = 0.0;
1621:     z[type == NORM_1_AND_2] = 0.0;
1622:   }
1623:   PetscFunctionReturn(PETSC_SUCCESS);
1624: }

1626: namespace detail
1627: {

1629: template <NormType wnormtype>
1630: class ErrorWNormTransformBase {
1631: public:
1632:   using result_type = thrust::tuple<PetscReal, PetscReal, PetscReal, PetscInt, PetscInt, PetscInt>;

1634:   constexpr explicit ErrorWNormTransformBase(PetscReal v) noexcept : ignore_max_{v} { }

1636: protected:
1637:   struct NormTuple {
1638:     PetscReal norm;
1639:     PetscInt  loc;
1640:   };

1642:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL static NormTuple compute_norm_(PetscReal err, PetscReal tol) noexcept
1643:   {
1644:     if (tol > 0.) {
1645:       const auto val = err / tol;

1647:       return {wnormtype == NORM_INFINITY ? val : PetscSqr(val), 1};
1648:     } else {
1649:       return {0.0, 0};
1650:     }
1651:   }

1653:   PetscReal ignore_max_;
1654: };

1656: template <NormType wnormtype>
1657: struct ErrorWNormTransform : ErrorWNormTransformBase<wnormtype> {
1658:   using base_type     = ErrorWNormTransformBase<wnormtype>;
1659:   using result_type   = typename base_type::result_type;
1660:   using argument_type = thrust::tuple<PetscScalar, PetscScalar, PetscScalar, PetscScalar>;

1662:   using base_type::base_type;

1664:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL result_type operator()(const argument_type &x) const noexcept
1665:   {
1666:     const auto u     = thrust::get<0>(x); // with x.get<0>(), cuda-12.4.0 gives error: class "cuda::std::__4::tuple<PetscScalar, PetscScalar, PetscScalar, PetscScalar>" has no member "get"
1667:     const auto y     = thrust::get<1>(x);
1668:     const auto au    = PetscAbsScalar(u);
1669:     const auto ay    = PetscAbsScalar(y);
1670:     const auto skip  = au < this->ignore_max_ || ay < this->ignore_max_;
1671:     const auto tola  = skip ? 0.0 : PetscRealPart(thrust::get<2>(x));
1672:     const auto tolr  = skip ? 0.0 : PetscRealPart(thrust::get<3>(x)) * PetscMax(au, ay);
1673:     const auto tol   = tola + tolr;
1674:     const auto err   = PetscAbsScalar(u - y);
1675:     const auto tup_a = this->compute_norm_(err, tola);
1676:     const auto tup_r = this->compute_norm_(err, tolr);
1677:     const auto tup_n = this->compute_norm_(err, tol);

1679:     return {tup_n.norm, tup_a.norm, tup_r.norm, tup_n.loc, tup_a.loc, tup_r.loc};
1680:   }
1681: };

1683: template <NormType wnormtype>
1684: struct ErrorWNormETransform : ErrorWNormTransformBase<wnormtype> {
1685:   using base_type     = ErrorWNormTransformBase<wnormtype>;
1686:   using result_type   = typename base_type::result_type;
1687:   using argument_type = thrust::tuple<PetscScalar, PetscScalar, PetscScalar, PetscScalar, PetscScalar>;

1689:   using base_type::base_type;

1691:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL result_type operator()(const argument_type &x) const noexcept
1692:   {
1693:     const auto au    = PetscAbsScalar(thrust::get<0>(x));
1694:     const auto ay    = PetscAbsScalar(thrust::get<1>(x));
1695:     const auto skip  = au < this->ignore_max_ || ay < this->ignore_max_;
1696:     const auto tola  = skip ? 0.0 : PetscRealPart(thrust::get<3>(x));
1697:     const auto tolr  = skip ? 0.0 : PetscRealPart(thrust::get<4>(x)) * PetscMax(au, ay);
1698:     const auto tol   = tola + tolr;
1699:     const auto err   = PetscAbsScalar(thrust::get<2>(x));
1700:     const auto tup_a = this->compute_norm_(err, tola);
1701:     const auto tup_r = this->compute_norm_(err, tolr);
1702:     const auto tup_n = this->compute_norm_(err, tol);

1704:     return {tup_n.norm, tup_a.norm, tup_r.norm, tup_n.loc, tup_a.loc, tup_r.loc};
1705:   }
1706: };

1708: template <NormType wnormtype>
1709: struct ErrorWNormReduce {
1710:   using value_type = typename ErrorWNormTransformBase<wnormtype>::result_type;

1712:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL value_type operator()(const value_type &lhs, const value_type &rhs) const noexcept
1713:   {
1714:     // cannot use lhs.get<0>() etc since the using decl above ambiguates the fact that
1715:     // result_type is a template, so in order to fix this we would need to write:
1716:     //
1717:     // lhs.template get<0>()
1718:     //
1719:     // which is unseemly.
1720:     if (wnormtype == NORM_INFINITY) {
1721:       // clang-format off
1722:       return {
1723:         PetscMax(thrust::get<0>(lhs), thrust::get<0>(rhs)),
1724:         PetscMax(thrust::get<1>(lhs), thrust::get<1>(rhs)),
1725:         PetscMax(thrust::get<2>(lhs), thrust::get<2>(rhs)),
1726:         thrust::get<3>(lhs) + thrust::get<3>(rhs),
1727:         thrust::get<4>(lhs) + thrust::get<4>(rhs),
1728:         thrust::get<5>(lhs) + thrust::get<5>(rhs)
1729:       };
1730:       // clang-format on
1731:     } else {
1732:       // clang-format off
1733:       return {
1734:         thrust::get<0>(lhs) + thrust::get<0>(rhs),
1735:         thrust::get<1>(lhs) + thrust::get<1>(rhs),
1736:         thrust::get<2>(lhs) + thrust::get<2>(rhs),
1737:         thrust::get<3>(lhs) + thrust::get<3>(rhs),
1738:         thrust::get<4>(lhs) + thrust::get<4>(rhs),
1739:         thrust::get<5>(lhs) + thrust::get<5>(rhs)
1740:       };
1741:       // clang-format on
1742:     }
1743:   }
1744: };

1746: template <template <NormType> class WNormTransformType, typename Tuple, typename cupmStream_t>
1747: inline PetscErrorCode ExecuteWNorm(Tuple &&first, Tuple &&last, NormType wnormtype, cupmStream_t stream, PetscReal ignore_max, PetscReal *norm, PetscInt *norm_loc, PetscReal *norma, PetscInt *norma_loc, PetscReal *normr, PetscInt *normr_loc) noexcept
1748: {
1749:   auto      begin = thrust::make_zip_iterator(std::forward<Tuple>(first));
1750:   auto      end   = thrust::make_zip_iterator(std::forward<Tuple>(last));
1751:   PetscReal n = 0, na = 0, nr = 0;
1752:   PetscInt  n_loc = 0, na_loc = 0, nr_loc = 0;

1754:   PetscFunctionBegin;
1755:   // clang-format off
1756:   if (wnormtype == NORM_INFINITY) {
1757:     PetscCallThrust(
1758:       thrust::tie(*norm, *norma, *normr, *norm_loc, *norma_loc, *normr_loc) = THRUST_CALL(
1759:         thrust::transform_reduce,
1760:         stream,
1761:         std::move(begin),
1762:         std::move(end),
1763:         WNormTransformType<NORM_INFINITY>{ignore_max},
1764:         thrust::make_tuple(n, na, nr, n_loc, na_loc, nr_loc),
1765:         ErrorWNormReduce<NORM_INFINITY>{}
1766:       )
1767:     );
1768:   } else {
1769:     PetscCallThrust(
1770:       thrust::tie(*norm, *norma, *normr, *norm_loc, *norma_loc, *normr_loc) = THRUST_CALL(
1771:         thrust::transform_reduce,
1772:         stream,
1773:         std::move(begin),
1774:         std::move(end),
1775:         WNormTransformType<NORM_2>{ignore_max},
1776:         thrust::make_tuple(n, na, nr, n_loc, na_loc, nr_loc),
1777:         ErrorWNormReduce<NORM_2>{}
1778:       )
1779:     );
1780:   }
1781:   // clang-format on
1782:   if (wnormtype == NORM_2) {
1783:     *norm  = PetscSqrtReal(*norm);
1784:     *norma = PetscSqrtReal(*norma);
1785:     *normr = PetscSqrtReal(*normr);
1786:   }
1787:   PetscFunctionReturn(PETSC_SUCCESS);
1788: }

1790: } // namespace detail

1792: // v->ops->errorwnorm
1793: template <device::cupm::DeviceType T>
1794: inline PetscErrorCode VecSeq_CUPM<T>::ErrorWnorm(Vec U, Vec Y, Vec E, NormType wnormtype, PetscReal atol, Vec vatol, PetscReal rtol, Vec vrtol, PetscReal ignore_max, PetscReal *norm, PetscInt *norm_loc, PetscReal *norma, PetscInt *norma_loc, PetscReal *normr, PetscInt *normr_loc) noexcept
1795: {
1796:   const auto         nl  = U->map->n;
1797:   auto               ait = thrust::make_constant_iterator(static_cast<PetscScalar>(atol));
1798:   auto               rit = thrust::make_constant_iterator(static_cast<PetscScalar>(rtol));
1799:   PetscDeviceContext dctx;
1800:   cupmStream_t       stream;

1802:   PetscFunctionBegin;
1803:   PetscCall(GetHandles_(&dctx, &stream));
1804:   {
1805:     const auto ConditionalDeviceArrayRead = [&](Vec v) {
1806:       if (v) {
1807:         return thrust::device_pointer_cast(DeviceArrayRead(dctx, v).data());
1808:       } else {
1809:         return thrust::device_ptr<PetscScalar>{nullptr};
1810:       }
1811:     };

1813:     const auto uarr = DeviceArrayRead(dctx, U);
1814:     const auto yarr = DeviceArrayRead(dctx, Y);
1815:     const auto uptr = thrust::device_pointer_cast(uarr.data());
1816:     const auto yptr = thrust::device_pointer_cast(yarr.data());
1817:     const auto eptr = ConditionalDeviceArrayRead(E);
1818:     const auto rptr = ConditionalDeviceArrayRead(vrtol);
1819:     const auto aptr = ConditionalDeviceArrayRead(vatol);

1821:     if (!vatol && !vrtol) {
1822:       if (E) {
1823:         // clang-format off
1824:         PetscCall(
1825:           detail::ExecuteWNorm<detail::ErrorWNormETransform>(
1826:             thrust::make_tuple(uptr, yptr, eptr, ait, rit),
1827:             thrust::make_tuple(uptr + nl, yptr + nl, eptr + nl, ait, rit),
1828:             wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1829:           )
1830:         );
1831:         // clang-format on
1832:       } else {
1833:         // clang-format off
1834:         PetscCall(
1835:           detail::ExecuteWNorm<detail::ErrorWNormTransform>(
1836:             thrust::make_tuple(uptr, yptr, ait, rit),
1837:             thrust::make_tuple(uptr + nl, yptr + nl, ait, rit),
1838:             wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1839:           )
1840:         );
1841:         // clang-format on
1842:       }
1843:     } else if (!vatol) {
1844:       if (E) {
1845:         // clang-format off
1846:         PetscCall(
1847:           detail::ExecuteWNorm<detail::ErrorWNormETransform>(
1848:             thrust::make_tuple(uptr, yptr, eptr, ait, rptr),
1849:             thrust::make_tuple(uptr + nl, yptr + nl, eptr + nl, ait, rptr + nl),
1850:             wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1851:           )
1852:         );
1853:         // clang-format on
1854:       } else {
1855:         // clang-format off
1856:         PetscCall(
1857:           detail::ExecuteWNorm<detail::ErrorWNormTransform>(
1858:             thrust::make_tuple(uptr, yptr, ait, rptr),
1859:             thrust::make_tuple(uptr + nl, yptr + nl, ait, rptr + nl),
1860:             wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1861:           )
1862:         );
1863:         // clang-format on
1864:       }
1865:     } else if (!vrtol) {
1866:       if (E) {
1867:         // clang-format off
1868:           PetscCall(
1869:             detail::ExecuteWNorm<detail::ErrorWNormETransform>(
1870:               thrust::make_tuple(uptr, yptr, eptr, aptr, rit),
1871:               thrust::make_tuple(uptr + nl, yptr + nl, eptr + nl, aptr + nl, rit),
1872:               wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1873:             )
1874:           );
1875:         // clang-format on
1876:       } else {
1877:         // clang-format off
1878:           PetscCall(
1879:             detail::ExecuteWNorm<detail::ErrorWNormTransform>(
1880:               thrust::make_tuple(uptr, yptr, aptr, rit),
1881:               thrust::make_tuple(uptr + nl, yptr + nl, aptr + nl, rit),
1882:               wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1883:             )
1884:           );
1885:         // clang-format on
1886:       }
1887:     } else {
1888:       if (E) {
1889:         // clang-format off
1890:           PetscCall(
1891:             detail::ExecuteWNorm<detail::ErrorWNormETransform>(
1892:               thrust::make_tuple(uptr, yptr, eptr, aptr, rptr),
1893:               thrust::make_tuple(uptr + nl, yptr + nl, eptr + nl, aptr + nl, rptr + nl),
1894:               wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1895:             )
1896:           );
1897:         // clang-format on
1898:       } else {
1899:         // clang-format off
1900:           PetscCall(
1901:             detail::ExecuteWNorm<detail::ErrorWNormTransform>(
1902:               thrust::make_tuple(uptr, yptr, aptr, rptr),
1903:               thrust::make_tuple(uptr + nl, yptr + nl, aptr + nl, rptr + nl),
1904:               wnormtype, stream, ignore_max, norm, norm_loc, norma, norma_loc, normr, normr_loc
1905:             )
1906:           );
1907:         // clang-format on
1908:       }
1909:     }
1910:   }
1911:   PetscFunctionReturn(PETSC_SUCCESS);
1912: }

1914: namespace detail
1915: {
1916: struct dotnorm2_mult {
1917:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL thrust::tuple<PetscScalar, PetscScalar> operator()(const PetscScalar &s, const PetscScalar &t) const noexcept
1918:   {
1919:     const auto conjt = PetscConj(t);

1921:     return {s * conjt, t * conjt};
1922:   }
1923: };

1925: // it is positively __bananas__ that thrust does not define default operator+ for tuples... I
1926: // would do it myself but now I am worried that they do so on purpose...
1927: struct dotnorm2_tuple_plus {
1928:   using value_type = thrust::tuple<PetscScalar, PetscScalar>;

1930:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL value_type operator()(const value_type &lhs, const value_type &rhs) const noexcept { return {thrust::get<0>(lhs) + thrust::get<0>(rhs), thrust::get<1>(lhs) + thrust::get<1>(rhs)}; }
1931: };

1933: } // namespace detail

1935: // v->ops->dotnorm2
1936: template <device::cupm::DeviceType T>
1937: inline PetscErrorCode VecSeq_CUPM<T>::DotNorm2(Vec s, Vec t, PetscScalar *dp, PetscScalar *nm) noexcept
1938: {
1939:   PetscDeviceContext dctx;
1940:   cupmStream_t       stream;

1942:   PetscFunctionBegin;
1943:   PetscCall(GetHandles_(&dctx, &stream));
1944:   {
1945:     PetscScalar dpt = 0.0, nmt = 0.0;
1946:     const auto  sdptr = thrust::device_pointer_cast(DeviceArrayRead(dctx, s).data());

1948:     // clang-format off
1949:     PetscCallThrust(
1950:       thrust::tie(*dp, *nm) = THRUST_CALL(
1951:         thrust::inner_product,
1952:         stream,
1953:         sdptr, sdptr+s->map->n, thrust::device_pointer_cast(DeviceArrayRead(dctx, t).data()),
1954:         thrust::make_tuple(dpt, nmt),
1955:         detail::dotnorm2_tuple_plus{}, detail::dotnorm2_mult{}
1956:       );
1957:     );
1958:     // clang-format on
1959:   }
1960:   PetscFunctionReturn(PETSC_SUCCESS);
1961: }

1963: namespace detail
1964: {
1965: struct conjugate {
1966:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL PetscScalar operator()(const PetscScalar &x) const noexcept { return PetscConj(x); }
1967: };

1969: } // namespace detail

1971: // v->ops->conjugate
1972: template <device::cupm::DeviceType T>
1973: inline PetscErrorCode VecSeq_CUPM<T>::ConjugateAsync(Vec xin, PetscDeviceContext dctx) noexcept
1974: {
1975:   PetscFunctionBegin;
1976:   if (PetscDefined(USE_COMPLEX)) PetscCall(PointwiseUnary_(detail::conjugate{}, xin, nullptr, dctx));
1977:   PetscFunctionReturn(PETSC_SUCCESS);
1978: }

1980: // v->ops->conjugate
1981: template <device::cupm::DeviceType T>
1982: inline PetscErrorCode VecSeq_CUPM<T>::Conjugate(Vec xin) noexcept
1983: {
1984:   PetscFunctionBegin;
1985:   PetscCall(ConjugateAsync(xin, nullptr));
1986:   PetscFunctionReturn(PETSC_SUCCESS);
1987: }

1989: namespace detail
1990: {

1992: struct real_part {
1993:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL thrust::tuple<PetscReal, PetscInt> operator()(const thrust::tuple<PetscScalar, PetscInt> &x) const noexcept { return {PetscRealPart(thrust::get<0>(x)), thrust::get<1>(x)}; }

1995:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL PetscReal operator()(const PetscScalar &x) const noexcept { return PetscRealPart(x); }
1996: };

1998: // deriving from Operator allows us to "store" an instance of the operator in the class but
1999: // also take advantage of empty base class optimization if the operator is stateless
2000: template <typename Operator>
2001: class tuple_compare : Operator {
2002: public:
2003:   using tuple_type    = thrust::tuple<PetscReal, PetscInt>;
2004:   using operator_type = Operator;

2006:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL tuple_type operator()(const tuple_type &x, const tuple_type &y) const noexcept
2007:   {
2008:     if (op_()(thrust::get<0>(y), thrust::get<0>(x))) {
2009:       // if y is strictly greater/less than x, return y
2010:       return y;
2011:     } else if (thrust::get<0>(y) == thrust::get<0>(x)) {
2012:       // if equal, prefer lower index
2013:       return thrust::get<1>(y) < thrust::get<1>(x) ? y : x;
2014:     }
2015:     // otherwise return x
2016:     return x;
2017:   }

2019: private:
2020:   PETSC_NODISCARD PETSC_HOSTDEVICE_INLINE_DECL const operator_type &op_() const noexcept { return *this; }
2021: };

2023: } // namespace detail

2025: template <device::cupm::DeviceType T>
2026: template <typename TupleFuncT, typename UnaryFuncT>
2027: inline PetscErrorCode VecSeq_CUPM<T>::MinMax_(TupleFuncT &&tuple_ftr, UnaryFuncT &&unary_ftr, Vec v, PetscInt *p, PetscReal *m) noexcept
2028: {
2029:   PetscFunctionBegin;
2030:   PetscCheckTypeNames(v, VECSEQCUPM(), VECMPICUPM());
2031:   if (p) *p = -1;
2032:   if (const auto n = v->map->n) {
2033:     PetscDeviceContext dctx;
2034:     cupmStream_t       stream;

2036:     PetscCall(GetHandles_(&dctx, &stream));
2037:     // needed to:
2038:     // 1. switch between transform_reduce and reduce
2039:     // 2. strip the real_part functor from the arguments
2040: #if PetscDefined(USE_COMPLEX)
2041:   #define THRUST_MINMAX_REDUCE(...) THRUST_CALL(thrust::transform_reduce, __VA_ARGS__)
2042: #else
2043:   #define THRUST_MINMAX_REDUCE(s, b, e, real_part__, ...) THRUST_CALL(thrust::reduce, s, b, e, __VA_ARGS__)
2044: #endif
2045:     {
2046:       const auto vptr = thrust::device_pointer_cast(DeviceArrayRead(dctx, v).data());

2048:       if (p) {
2049:         // clang-format off
2050:         const auto zip = thrust::make_zip_iterator(
2051:           thrust::make_tuple(std::move(vptr), thrust::make_counting_iterator(PetscInt{0}))
2052:         );
2053:         // clang-format on
2054:         // need to use preprocessor conditionals since otherwise thrust complains about not being
2055:         // able to convert a thrust::device_reference<PetscScalar> to a PetscReal on complex
2056:         // builds...
2057:         // clang-format off
2058:         PetscCallThrust(
2059:           thrust::tie(*m, *p) = THRUST_MINMAX_REDUCE(
2060:             stream, zip, zip + n, detail::real_part{},
2061:             thrust::make_tuple(*m, *p), std::forward<TupleFuncT>(tuple_ftr)
2062:           );
2063:         );
2064:         // clang-format on
2065:       } else {
2066:         // clang-format off
2067:         PetscCallThrust(
2068:           *m = THRUST_MINMAX_REDUCE(
2069:             stream, vptr, vptr + n, detail::real_part{},
2070:             *m, std::forward<UnaryFuncT>(unary_ftr)
2071:           );
2072:         );
2073:         // clang-format on
2074:       }
2075:     }
2076: #undef THRUST_MINMAX_REDUCE
2077:   }
2078:   // REVIEW ME: flops?
2079:   PetscFunctionReturn(PETSC_SUCCESS);
2080: }

2082: // v->ops->max
2083: template <device::cupm::DeviceType T>
2084: inline PetscErrorCode VecSeq_CUPM<T>::Max(Vec v, PetscInt *p, PetscReal *m) noexcept
2085: {
2086:   using tuple_functor = detail::tuple_compare<thrust::greater<PetscReal>>;
2087:   using unary_functor = thrust::maximum<PetscReal>;

2089:   PetscFunctionBegin;
2090:   *m = PETSC_MIN_REAL;
2091:   // use {} constructor syntax otherwise most vexing parse
2092:   PetscCall(MinMax_(tuple_functor{}, unary_functor{}, v, p, m));
2093:   PetscFunctionReturn(PETSC_SUCCESS);
2094: }

2096: // v->ops->min
2097: template <device::cupm::DeviceType T>
2098: inline PetscErrorCode VecSeq_CUPM<T>::Min(Vec v, PetscInt *p, PetscReal *m) noexcept
2099: {
2100:   using tuple_functor = detail::tuple_compare<thrust::less<PetscReal>>;
2101:   using unary_functor = thrust::minimum<PetscReal>;

2103:   PetscFunctionBegin;
2104:   *m = PETSC_MAX_REAL;
2105:   // use {} constructor syntax otherwise most vexing parse
2106:   PetscCall(MinMax_(tuple_functor{}, unary_functor{}, v, p, m));
2107:   PetscFunctionReturn(PETSC_SUCCESS);
2108: }

2110: // v->ops->sum
2111: template <device::cupm::DeviceType T>
2112: inline PetscErrorCode VecSeq_CUPM<T>::Sum(Vec v, PetscScalar *sum) noexcept
2113: {
2114:   PetscFunctionBegin;
2115:   if (const auto n = v->map->n) {
2116:     PetscDeviceContext dctx;
2117:     cupmStream_t       stream;

2119:     PetscCall(GetHandles_(&dctx, &stream));
2120:     const auto dptr = thrust::device_pointer_cast(DeviceArrayRead(dctx, v).data());
2121:     // REVIEW ME: why not cupmBlasXasum()?
2122:     PetscCallThrust(*sum = THRUST_CALL(thrust::reduce, stream, dptr, dptr + n, PetscScalar{0.0}););
2123:     // REVIEW ME: must be at least n additions
2124:     PetscCall(PetscLogGpuFlops(n));
2125:   } else {
2126:     *sum = 0.0;
2127:   }
2128:   PetscFunctionReturn(PETSC_SUCCESS);
2129: }

2131: template <device::cupm::DeviceType T>
2132: inline PetscErrorCode VecSeq_CUPM<T>::ShiftAsync(Vec v, PetscScalar shift, PetscDeviceContext dctx) noexcept
2133: {
2134:   PetscFunctionBegin;
2135:   PetscCall(PointwiseUnary_(device::cupm::functors::make_plus_equals(shift), v, nullptr, dctx));
2136:   PetscFunctionReturn(PETSC_SUCCESS);
2137: }

2139: template <device::cupm::DeviceType T>
2140: inline PetscErrorCode VecSeq_CUPM<T>::Shift(Vec v, PetscScalar shift) noexcept
2141: {
2142:   PetscFunctionBegin;
2143:   PetscCall(ShiftAsync(v, shift, nullptr));
2144:   PetscFunctionReturn(PETSC_SUCCESS);
2145: }

2147: template <device::cupm::DeviceType T>
2148: inline PetscErrorCode VecSeq_CUPM<T>::SetRandom(Vec v, PetscRandom rand) noexcept
2149: {
2150:   PetscFunctionBegin;
2151:   if (const auto n = v->map->n) {
2152:     PetscBool          iscurand;
2153:     PetscDeviceContext dctx;

2155:     PetscCall(GetHandles_(&dctx));
2156:     PetscCall(PetscObjectTypeCompare(PetscObjectCast(rand), PETSCCURAND, &iscurand));
2157:     if (iscurand) PetscCall(PetscRandomGetValues(rand, n, DeviceArrayWrite(dctx, v)));
2158:     else PetscCall(PetscRandomGetValues(rand, n, HostArrayWrite(dctx, v)));
2159:   } else {
2160:     PetscCall(MaybeIncrementEmptyLocalVec(v));
2161:   }
2162:   // REVIEW ME: flops????
2163:   // REVIEW ME: Timing???
2164:   PetscFunctionReturn(PETSC_SUCCESS);
2165: }

2167: // v->ops->setpreallocation
2168: template <device::cupm::DeviceType T>
2169: inline PetscErrorCode VecSeq_CUPM<T>::SetPreallocationCOO(Vec v, PetscCount ncoo, const PetscInt coo_i[]) noexcept
2170: {
2171:   PetscDeviceContext dctx;

2173:   PetscFunctionBegin;
2174:   PetscCall(GetHandles_(&dctx));
2175:   PetscCall(VecSetPreallocationCOO_Seq(v, ncoo, coo_i));
2176:   PetscCall(SetPreallocationCOO_CUPMBase(v, ncoo, coo_i, dctx));
2177:   PetscFunctionReturn(PETSC_SUCCESS);
2178: }

2180: // v->ops->setvaluescoo
2181: template <device::cupm::DeviceType T>
2182: inline PetscErrorCode VecSeq_CUPM<T>::SetValuesCOO(Vec x, const PetscScalar v[], InsertMode imode) noexcept
2183: {
2184:   auto               vv = const_cast<PetscScalar *>(v);
2185:   PetscMemType       memtype;
2186:   PetscDeviceContext dctx;
2187:   cupmStream_t       stream;

2189:   PetscFunctionBegin;
2190:   PetscCall(GetHandles_(&dctx, &stream));
2191:   PetscCall(PetscGetMemType(v, &memtype));
2192:   if (PetscMemTypeHost(memtype)) {
2193:     const auto size = VecIMPLCast(x)->coo_n;

2195:     // If user gave v[] in host, we might need to copy it to device if any
2196:     PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_CUPM(), size, &vv));
2197:     PetscCall(PetscCUPMMemcpyAsync(vv, v, size, cupmMemcpyHostToDevice, stream));
2198:   }

2200:   if (const auto n = x->map->n) {
2201:     const auto vcu = VecCUPMCast(x);

2203:     PetscCall(PetscCUPMLaunchKernel1D(n, 0, stream, kernels::add_coo_values, vv, n, vcu->jmap1_d, vcu->perm1_d, imode, imode == INSERT_VALUES ? DeviceArrayWrite(dctx, x).data() : DeviceArrayReadWrite(dctx, x).data()));
2204:   } else {
2205:     PetscCall(MaybeIncrementEmptyLocalVec(x));
2206:   }

2208:   if (PetscMemTypeHost(memtype)) PetscCall(PetscDeviceFree(dctx, vv));
2209:   PetscCall(PetscDeviceContextSynchronize(dctx));
2210:   PetscFunctionReturn(PETSC_SUCCESS);
2211: }

2213: } // namespace impl

2215: } // namespace cupm

2217: } // namespace vec

2219: } // namespace Petsc