Actual source code: sfcuda.cu

petsc-3.15.0 2021-03-30
Report Typos and Errors
  1: #include <../src/vec/is/sf/impls/basic/sfpack.h>
  2: #include <cuda_runtime.h>
  3: #include <petsccublas.h>

  5: /* Map a thread id to an index in root/leaf space through a series of 3D subdomains. See PetscSFPackOpt. */
  6: __device__ static inline PetscInt MapTidToIndex(const PetscInt *opt,PetscInt tid)
  7: {
  8:   PetscInt        i,j,k,m,n,r;
  9:   const PetscInt  *offset,*start,*dx,*dy,*X,*Y;

 11:   n      = opt[0];
 12:   offset = opt + 1;
 13:   start  = opt + n + 2;
 14:   dx     = opt + 2*n + 2;
 15:   dy     = opt + 3*n + 2;
 16:   X      = opt + 5*n + 2;
 17:   Y      = opt + 6*n + 2;
 18:   for (r=0; r<n; r++) {if (tid < offset[r+1]) break;}
 19:   m = (tid - offset[r]);
 20:   k = m/(dx[r]*dy[r]);
 21:   j = (m - k*dx[r]*dy[r])/dx[r];
 22:   i = m - k*dx[r]*dy[r] - j*dx[r];

 24:   return (start[r] + k*X[r]*Y[r] + j*X[r] + i);
 25: }

 27: /*====================================================================================*/
 28: /*  Templated CUDA kernels for pack/unpack. The Op can be regular or atomic           */
 29: /*====================================================================================*/

 31: /* Suppose user calls PetscSFReduce(sf,unit,...) and <unit> is an MPI data type made of 16 PetscReals, then
 32:    <Type> is PetscReal, which is the primitive type we operate on.
 33:    <bs>   is 16, which says <unit> contains 16 primitive types.
 34:    <BS>   is 8, which is the maximal SIMD width we will try to vectorize operations on <unit>.
 35:    <EQ>   is 0, which is (bs == BS ? 1 : 0)

 37:   If instead, <unit> has 8 PetscReals, then bs=8, BS=8, EQ=1, rendering MBS below to a compile time constant.
 38:   For the common case in VecScatter, bs=1, BS=1, EQ=1, MBS=1, the inner for-loops below will be totally unrolled.
 39: */
 40: template<class Type,PetscInt BS,PetscInt EQ>
 41: __global__ static void d_Pack(PetscInt bs,PetscInt count,PetscInt start,const PetscInt *opt,const PetscInt *idx,const Type *data,Type *buf)
 42: {
 43:   PetscInt        i,s,t,tid = blockIdx.x*blockDim.x + threadIdx.x;
 44:   const PetscInt  grid_size = gridDim.x * blockDim.x;
 45:   const PetscInt  M = (EQ) ? 1 : bs/BS; /* If EQ, then M=1 enables compiler's const-propagation */
 46:   const PetscInt  MBS = M*BS;  /* MBS=bs. We turn MBS into a compile-time const when EQ=1. */

 48:   for (; tid<count; tid += grid_size) {
 49:     /* opt != NULL ==> idx == NULL, i.e., the indices have patterns but not contiguous;
 50:        opt == NULL && idx == NULL ==> the indices are contiguous;
 51:      */
 52:     t = (opt? MapTidToIndex(opt,tid) : (idx? idx[tid] : start+tid))*MBS;
 53:     s = tid*MBS;
 54:     for (i=0; i<MBS; i++) buf[s+i] = data[t+i];
 55:   }
 56: }

 58: template<class Type,class Op,PetscInt BS,PetscInt EQ>
 59: __global__ static void d_UnpackAndOp(PetscInt bs,PetscInt count,PetscInt start,const PetscInt *opt,const PetscInt *idx,Type *data,const Type *buf)
 60: {
 61:   PetscInt        i,s,t,tid = blockIdx.x*blockDim.x + threadIdx.x;
 62:   const PetscInt  grid_size = gridDim.x * blockDim.x;
 63:   const PetscInt  M = (EQ) ? 1 : bs/BS, MBS = M*BS;
 64:   Op              op;

 66:   for (; tid<count; tid += grid_size) {
 67:     t = (opt? MapTidToIndex(opt,tid) : (idx? idx[tid] : start+tid))*MBS;
 68:     s = tid*MBS;
 69:     for (i=0; i<MBS; i++) op(data[t+i],buf[s+i]);
 70:   }
 71: }

 73: template<class Type,class Op,PetscInt BS,PetscInt EQ>
 74: __global__ static void d_FetchAndOp(PetscInt bs,PetscInt count,PetscInt rootstart,const PetscInt *rootopt,const PetscInt *rootidx,Type *rootdata,Type *leafbuf)
 75: {
 76:   PetscInt        i,l,r,tid = blockIdx.x*blockDim.x + threadIdx.x;
 77:   const PetscInt  grid_size = gridDim.x * blockDim.x;
 78:   const PetscInt  M = (EQ) ? 1 : bs/BS, MBS = M*BS;
 79:   Op              op;

 81:   for (; tid<count; tid += grid_size) {
 82:     r = (rootopt? MapTidToIndex(rootopt,tid) : (rootidx? rootidx[tid] : rootstart+tid))*MBS;
 83:     l = tid*MBS;
 84:     for (i=0; i<MBS; i++) leafbuf[l+i] = op(rootdata[r+i],leafbuf[l+i]);
 85:   }
 86: }

 88: template<class Type,class Op,PetscInt BS,PetscInt EQ>
 89: __global__ static void d_ScatterAndOp(PetscInt bs,PetscInt count,PetscInt srcx,PetscInt srcy,PetscInt srcX,PetscInt srcY,PetscInt srcStart,const PetscInt* srcIdx,const Type *src,PetscInt dstx,PetscInt dsty,PetscInt dstX,PetscInt dstY,PetscInt dstStart,const PetscInt *dstIdx,Type *dst)
 90: {
 91:   PetscInt        i,j,k,s,t,tid = blockIdx.x*blockDim.x + threadIdx.x;
 92:   const PetscInt  grid_size = gridDim.x * blockDim.x;
 93:   const PetscInt  M = (EQ) ? 1 : bs/BS, MBS = M*BS;
 94:   Op              op;

 96:   for (; tid<count; tid += grid_size) {
 97:     if (!srcIdx) { /* src is either contiguous or 3D */
 98:       k = tid/(srcx*srcy);
 99:       j = (tid - k*srcx*srcy)/srcx;
100:       i = tid - k*srcx*srcy - j*srcx;
101:       s = srcStart + k*srcX*srcY + j*srcX + i;
102:     } else {
103:       s = srcIdx[tid];
104:     }

106:     if (!dstIdx) { /* dst is either contiguous or 3D */
107:       k = tid/(dstx*dsty);
108:       j = (tid - k*dstx*dsty)/dstx;
109:       i = tid - k*dstx*dsty - j*dstx;
110:       t = dstStart + k*dstX*dstY + j*dstX + i;
111:     } else {
112:       t = dstIdx[tid];
113:     }

115:     s *= MBS;
116:     t *= MBS;
117:     for (i=0; i<MBS; i++) op(dst[t+i],src[s+i]);
118:   }
119: }

121: template<class Type,class Op,PetscInt BS,PetscInt EQ>
122: __global__ static void d_FetchAndOpLocal(PetscInt bs,PetscInt count,PetscInt rootstart,const PetscInt *rootopt,const PetscInt *rootidx,Type *rootdata,PetscInt leafstart,const PetscInt *leafopt,const PetscInt *leafidx,const Type *leafdata,Type *leafupdate)
123: {
124:   PetscInt        i,l,r,tid = blockIdx.x*blockDim.x + threadIdx.x;
125:   const PetscInt  grid_size = gridDim.x * blockDim.x;
126:   const PetscInt  M = (EQ) ? 1 : bs/BS, MBS = M*BS;
127:   Op              op;

129:   for (; tid<count; tid += grid_size) {
130:     r = (rootopt? MapTidToIndex(rootopt,tid) : (rootidx? rootidx[tid] : rootstart+tid))*MBS;
131:     l = (leafopt? MapTidToIndex(leafopt,tid) : (leafidx? leafidx[tid] : leafstart+tid))*MBS;
132:     for (i=0; i<MBS; i++) leafupdate[l+i] = op(rootdata[r+i],leafdata[l+i]);
133:   }
134: }

136: /*====================================================================================*/
137: /*                             Regular operations on device                           */
138: /*====================================================================================*/
139: template<typename Type> struct Insert {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = y;             return old;}};
140: template<typename Type> struct Add    {__device__ Type operator() (Type& x,Type y) const {Type old = x; x += y;             return old;}};
141: template<typename Type> struct Mult   {__device__ Type operator() (Type& x,Type y) const {Type old = x; x *= y;             return old;}};
142: template<typename Type> struct Min    {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = PetscMin(x,y); return old;}};
143: template<typename Type> struct Max    {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = PetscMax(x,y); return old;}};
144: template<typename Type> struct LAND   {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = x && y;        return old;}};
145: template<typename Type> struct LOR    {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = x || y;        return old;}};
146: template<typename Type> struct LXOR   {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = !x != !y;      return old;}};
147: template<typename Type> struct BAND   {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = x & y;         return old;}};
148: template<typename Type> struct BOR    {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = x | y;         return old;}};
149: template<typename Type> struct BXOR   {__device__ Type operator() (Type& x,Type y) const {Type old = x; x  = x ^ y;         return old;}};
150: template<typename Type> struct Minloc {
151:   __device__ Type operator() (Type& x,Type y) const {
152:     Type old = x;
153:     if (y.a < x.a) x = y;
154:     else if (y.a == x.a) x.b = min(x.b,y.b);
155:     return old;
156:   }
157: };
158: template<typename Type> struct Maxloc {
159:   __device__ Type operator() (Type& x,Type y) const {
160:     Type old = x;
161:     if (y.a > x.a) x = y;
162:     else if (y.a == x.a) x.b = min(x.b,y.b); /* See MPI MAXLOC */
163:     return old;
164:   }
165: };

167: /*====================================================================================*/
168: /*                             Atomic operations on device                            */
169: /*====================================================================================*/

171: /*
172:   Atomic Insert (exchange) operations

174:   CUDA C Programming Guide V10.1 Chapter B.12.1.3:

176:   int atomicExch(int* address, int val);
177:   unsigned int atomicExch(unsigned int* address, unsigned int val);
178:   unsigned long long int atomicExch(unsigned long long int* address, unsigned long long int val);
179:   float atomicExch(float* address, float val);

181:   reads the 32-bit or 64-bit word old located at the address address in global or shared
182:   memory and stores val back to memory at the same address. These two operations are
183:   performed in one atomic transaction. The function returns old.

185:   PETSc notes:

187:   It may be useful in PetscSFFetchAndOp with op = MPI_REPLACE.

189:   VecScatter with multiple entries scattered to the same location using INSERT_VALUES does not need
190:   atomic insertion, since it does not need the old value. A 32-bit or 64-bit store instruction should
191:   be atomic itself.

193:   With bs>1 and a unit > 64 bits, the current element-wise atomic approach can not guarantee the whole
194:   insertion is atomic. Hope no user codes rely on that.
195: */
196: __device__ static double atomicExch(double* address,double val) {return __longlong_as_double(atomicExch((ullint*)address,__double_as_longlong(val)));}

198: __device__ static llint atomicExch(llint* address,llint val) {return (llint)(atomicExch((ullint*)address,(ullint)val));}

200: template<typename Type> struct AtomicInsert {__device__ Type operator() (Type& x,Type y) const {return atomicExch(&x,y);}};

202: #if defined(PETSC_HAVE_COMPLEX)
203: #if defined(PETSC_USE_REAL_DOUBLE)
204: /* CUDA does not support 128-bit atomics. Users should not insert different 128-bit PetscComplex values to the same location */
205: template<> struct AtomicInsert<PetscComplex> {
206:   __device__ PetscComplex operator() (PetscComplex& x,PetscComplex y) const {
207:     PetscComplex         old, *z = &old;
208:     double               *xp = (double*)&x,*yp = (double*)&y;
209:     AtomicInsert<double> op;
210:     z[0] = op(xp[0],yp[0]);
211:     z[1] = op(xp[1],yp[1]);
212:     return old; /* The returned value may not be atomic. It can be mix of two ops. Caller should discard it. */
213:   }
214: };
215: #elif defined(PETSC_USE_REAL_SINGLE)
216: template<> struct AtomicInsert<PetscComplex> {
217:   __device__ PetscComplex operator() (PetscComplex& x,PetscComplex y) const {
218:     double               *xp = (double*)&x,*yp = (double*)&y;
219:     AtomicInsert<double> op;
220:     return op(xp[0],yp[0]);
221:   }
222: };
223: #endif
224: #endif

226: /*
227:   Atomic add operations

229:   CUDA C Programming Guide V10.1 Chapter B.12.1.1:

231:   int atomicAdd(int* address, int val);
232:   unsigned int atomicAdd(unsigned int* address,unsigned int val);
233:   unsigned long long int atomicAdd(unsigned long long int* address,unsigned long long int val);
234:   float atomicAdd(float* address, float val);
235:   double atomicAdd(double* address, double val);
236:   __half2 atomicAdd(__half2 *address, __half2 val);
237:   __half atomicAdd(__half *address, __half val);

239:   reads the 16-bit, 32-bit or 64-bit word old located at the address address in global or shared memory, computes (old + val),
240:   and stores the result back to memory at the same address. These three operations are performed in one atomic transaction. The
241:   function returns old.

243:   The 32-bit floating-point version of atomicAdd() is only supported by devices of compute capability 2.x and higher.
244:   The 64-bit floating-point version of atomicAdd() is only supported by devices of compute capability 6.x and higher.
245:   The 32-bit __half2 floating-point version of atomicAdd() is only supported by devices of compute capability 6.x and
246:   higher. The atomicity of the __half2 add operation is guaranteed separately for each of the two __half elements;
247:   the entire __half2 is not guaranteed to be atomic as a single 32-bit access.
248:   The 16-bit __half floating-point version of atomicAdd() is only supported by devices of compute capability 7.x and higher.
249: */
250: __device__ static llint atomicAdd(llint* address,llint val) {return (llint)atomicAdd((ullint*)address,(ullint)val);}

252: template<typename Type> struct AtomicAdd {__device__ Type operator() (Type& x,Type y) const {return atomicAdd(&x,y);}};

254: template<> struct AtomicAdd<double> {
255:   __device__ double operator() (double& x,double y) const {
256: #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 600)
257:     return atomicAdd(&x,y);
258: #else
259:     double *address = &x, val = y;
260:     ullint *address_as_ull = (ullint*)address;
261:     ullint old = *address_as_ull, assumed;
262:     do {
263:       assumed = old;
264:       old     = atomicCAS(address_as_ull, assumed, __double_as_longlong(val + __longlong_as_double(assumed)));
265:       /* Note: uses integer comparison to avoid hang in case of NaN (since NaN !=NaN) */
266:     } while (assumed != old);
267:     return __longlong_as_double(old);
268: #endif
269:   }
270: };

272: template<> struct AtomicAdd<float> {
273:   __device__ float operator() (float& x,float y) const {
274: #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
275:     return atomicAdd(&x,y);
276: #else
277:     float *address = &x, val = y;
278:     int   *address_as_int = (int*)address;
279:     int   old = *address_as_int, assumed;
280:     do {
281:       assumed = old;
282:       old     = atomicCAS(address_as_int, assumed, __float_as_int(val + __int_as_float(assumed)));
283:       /* Note: uses integer comparison to avoid hang in case of NaN (since NaN !=NaN) */
284:     } while (assumed != old);
285:     return __int_as_float(old);
286: #endif
287:   }
288: };

290: #if defined(PETSC_HAVE_COMPLEX)
291: template<> struct AtomicAdd<PetscComplex> {
292:  __device__ PetscComplex operator() (PetscComplex& x,PetscComplex y) const {
293:   PetscComplex         old, *z = &old;
294:   PetscReal            *xp = (PetscReal*)&x,*yp = (PetscReal*)&y;
295:   AtomicAdd<PetscReal> op;
296:   z[0] = op(xp[0],yp[0]);
297:   z[1] = op(xp[1],yp[1]);
298:   return old; /* The returned value may not be atomic. It can be mix of two ops. Caller should discard it. */
299:  }
300: };
301: #endif

303: /*
304:   Atomic Mult operations:

306:   CUDA has no atomicMult at all, so we build our own with atomicCAS
307:  */
308: #if defined(PETSC_USE_REAL_DOUBLE)
309: __device__ static double atomicMult(double* address, double val)
310: {
311:   ullint *address_as_ull = (ullint*)(address);
312:   ullint old = *address_as_ull, assumed;
313:   do {
314:     assumed = old;
315:     /* Other threads can access and modify value of *address_as_ull after the read above and before the write below */
316:     old     = atomicCAS(address_as_ull, assumed, __double_as_longlong(val*__longlong_as_double(assumed)));
317:   } while (assumed != old);
318:   return __longlong_as_double(old);
319: }
320: #elif defined(PETSC_USE_REAL_SINGLE)
321: __device__ static float atomicMult(float* address,float val)
322: {
323:   int *address_as_int = (int*)(address);
324:   int old = *address_as_int, assumed;
325:   do {
326:     assumed  = old;
327:     old      = atomicCAS(address_as_int, assumed, __float_as_int(val*__int_as_float(assumed)));
328:   } while (assumed != old);
329:   return __int_as_float(old);
330: }
331: #endif

333: __device__ static int atomicMult(int* address,int val)
334: {
335:   int *address_as_int = (int*)(address);
336:   int old = *address_as_int, assumed;
337:   do {
338:     assumed = old;
339:     old     = atomicCAS(address_as_int, assumed, val*assumed);
340:   } while (assumed != old);
341:   return (int)old;
342: }

344: __device__ static llint atomicMult(llint* address,llint val)
345: {
346:   ullint *address_as_ull = (ullint*)(address);
347:   ullint old = *address_as_ull, assumed;
348:   do {
349:     assumed = old;
350:     old     = atomicCAS(address_as_ull, assumed, (ullint)(val*(llint)assumed));
351:   } while (assumed != old);
352:   return (llint)old;
353: }

355: template<typename Type> struct AtomicMult {__device__ Type operator() (Type& x,Type y) const {return atomicMult(&x,y);}};

357: /*
358:   Atomic Min/Max operations

360:   CUDA C Programming Guide V10.1 Chapter B.12.1.4~5:

362:   int atomicMin(int* address, int val);
363:   unsigned int atomicMin(unsigned int* address,unsigned int val);
364:   unsigned long long int atomicMin(unsigned long long int* address,unsigned long long int val);

366:   reads the 32-bit or 64-bit word old located at the address address in global or shared
367:   memory, computes the minimum of old and val, and stores the result back to memory
368:   at the same address. These three operations are performed in one atomic transaction.
369:   The function returns old.
370:   The 64-bit version of atomicMin() is only supported by devices of compute capability 3.5 and higher.

372:   atomicMax() is similar.
373:  */

375: #if defined(PETSC_USE_REAL_DOUBLE)
376: __device__ static double atomicMin(double* address, double val)
377: {
378:   ullint *address_as_ull = (ullint*)(address);
379:   ullint old = *address_as_ull, assumed;
380:   do {
381:     assumed = old;
382:     old     = atomicCAS(address_as_ull, assumed, __double_as_longlong(PetscMin(val,__longlong_as_double(assumed))));
383:   } while (assumed != old);
384:   return __longlong_as_double(old);
385: }

387: __device__ static double atomicMax(double* address, double val)
388: {
389:   ullint *address_as_ull = (ullint*)(address);
390:   ullint old = *address_as_ull, assumed;
391:   do {
392:     assumed  = old;
393:     old = atomicCAS(address_as_ull, assumed, __double_as_longlong(PetscMax(val,__longlong_as_double(assumed))));
394:   } while (assumed != old);
395:   return __longlong_as_double(old);
396: }
397: #elif defined(PETSC_USE_REAL_SINGLE)
398: __device__ static float atomicMin(float* address,float val)
399: {
400:   int *address_as_int = (int*)(address);
401:   int old = *address_as_int, assumed;
402:   do {
403:     assumed = old;
404:     old     = atomicCAS(address_as_int, assumed, __float_as_int(PetscMin(val,__int_as_float(assumed))));
405:   } while (assumed != old);
406:   return __int_as_float(old);
407: }

409: __device__ static float atomicMax(float* address,float val)
410: {
411:   int *address_as_int = (int*)(address);
412:   int old = *address_as_int, assumed;
413:   do {
414:     assumed = old;
415:     old     = atomicCAS(address_as_int, assumed, __float_as_int(PetscMax(val,__int_as_float(assumed))));
416:   } while (assumed != old);
417:   return __int_as_float(old);
418: }
419: #endif

421: /*
422:   atomicMin/Max(long long *, long long) are not in Nvidia's documentation. But on OLCF Summit we found
423:   atomicMin/Max/And/Or/Xor(long long *, long long) in /sw/summit/cuda/10.1.243/include/sm_32_atomic_functions.h.
424:   This causes compilation errors with pgi compilers and 64-bit indices:
425:       error: function "atomicMin(long long *, long long)" has already been defined

427:   So we add extra conditions defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 320)
428: */
429: #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 320)
430: __device__ static llint atomicMin(llint* address,llint val)
431: {
432:   ullint *address_as_ull = (ullint*)(address);
433:   ullint old = *address_as_ull, assumed;
434:   do {
435:     assumed = old;
436:     old     = atomicCAS(address_as_ull, assumed, (ullint)(PetscMin(val,(llint)assumed)));
437:   } while (assumed != old);
438:   return (llint)old;
439: }

441: __device__ static llint atomicMax(llint* address,llint val)
442: {
443:   ullint *address_as_ull = (ullint*)(address);
444:   ullint old = *address_as_ull, assumed;
445:   do {
446:     assumed = old;
447:     old     = atomicCAS(address_as_ull, assumed, (ullint)(PetscMax(val,(llint)assumed)));
448:   } while (assumed != old);
449:   return (llint)old;
450: }
451: #endif

453: template<typename Type> struct AtomicMin {__device__ Type operator() (Type& x,Type y) const {return atomicMin(&x,y);}};
454: template<typename Type> struct AtomicMax {__device__ Type operator() (Type& x,Type y) const {return atomicMax(&x,y);}};

456: /*
457:   Atomic bitwise operations

459:   CUDA C Programming Guide V10.1 Chapter B.12.2.1 ~ B.12.2.3:

461:   int atomicAnd(int* address, int val);
462:   unsigned int atomicAnd(unsigned int* address,unsigned int val);
463:   unsigned long long int atomicAnd(unsigned long long int* address,unsigned long long int val);

465:   reads the 32-bit or 64-bit word old located at the address address in global or shared
466:   memory, computes (old & val), and stores the result back to memory at the same
467:   address. These three operations are performed in one atomic transaction.
468:   The function returns old.

470:   The 64-bit version of atomicAnd() is only supported by devices of compute capability 3.5 and higher.

472:   atomicOr() and atomicXor are similar.
473: */

475: #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 320) /* Why 320? see comments at atomicMin() above */
476: __device__ static llint atomicAnd(llint* address,llint val)
477: {
478:   ullint *address_as_ull = (ullint*)(address);
479:   ullint old = *address_as_ull, assumed;
480:   do {
481:     assumed = old;
482:     old     = atomicCAS(address_as_ull, assumed, (ullint)(val & (llint)assumed));
483:   } while (assumed != old);
484:   return (llint)old;
485: }
486: __device__ static llint atomicOr(llint* address,llint val)
487: {
488:   ullint *address_as_ull = (ullint*)(address);
489:   ullint old = *address_as_ull, assumed;
490:   do {
491:     assumed = old;
492:     old     = atomicCAS(address_as_ull, assumed, (ullint)(val | (llint)assumed));
493:   } while (assumed != old);
494:   return (llint)old;
495: }

497: __device__ static llint atomicXor(llint* address,llint val)
498: {
499:   ullint *address_as_ull = (ullint*)(address);
500:   ullint old = *address_as_ull, assumed;
501:   do {
502:     assumed = old;
503:     old     = atomicCAS(address_as_ull, assumed, (ullint)(val ^ (llint)assumed));
504:   } while (assumed != old);
505:   return (llint)old;
506: }
507: #endif

509: template<typename Type> struct AtomicBAND {__device__ Type operator() (Type& x,Type y) const {return atomicAnd(&x,y);}};
510: template<typename Type> struct AtomicBOR  {__device__ Type operator() (Type& x,Type y) const {return atomicOr (&x,y);}};
511: template<typename Type> struct AtomicBXOR {__device__ Type operator() (Type& x,Type y) const {return atomicXor(&x,y);}};

513: /*
514:   Atomic logical operations:

516:   CUDA has no atomic logical operations at all. We support them on integer types.
517: */

519: /* A template without definition makes any instantiation not using given specializations erroneous at compile time,
520:    which is what we want since we only support 32-bit and 64-bit integers.
521:  */
522: template<typename Type,class Op,int size/* sizeof(Type) */> struct AtomicLogical;

524: template<typename Type,class Op>
525: struct AtomicLogical<Type,Op,4> {
526:   __device__ Type operator()(Type& x,Type y) const {
527:     int *address_as_int = (int*)(&x);
528:     int old = *address_as_int, assumed;
529:     Op op;
530:     do {
531:       assumed = old;
532:       old     = atomicCAS(address_as_int, assumed, (int)(op((Type)assumed,y)));
533:     } while (assumed != old);
534:     return (Type)old;
535:   }
536: };

538: template<typename Type,class Op>
539: struct AtomicLogical<Type,Op,8> {
540:   __device__ Type operator()(Type& x,Type y) const {
541:     ullint *address_as_ull = (ullint*)(&x);
542:     ullint old = *address_as_ull, assumed;
543:     Op op;
544:     do {
545:       assumed = old;
546:       old     = atomicCAS(address_as_ull, assumed, (ullint)(op((Type)assumed,y)));
547:     } while (assumed != old);
548:     return (Type)old;
549:   }
550: };

552: /* Note land/lor/lxor below are different from LAND etc above. Here we pass arguments by value and return result of ops (not old value) */
553: template<typename Type> struct land {__device__ Type operator()(Type x, Type y) {return x && y;}};
554: template<typename Type> struct lor  {__device__ Type operator()(Type x, Type y) {return x || y;}};
555: template<typename Type> struct lxor {__device__ Type operator()(Type x, Type y) {return (!x != !y);}};

557: template<typename Type> struct AtomicLAND {__device__ Type operator()(Type& x,Type y) const {AtomicLogical<Type,land<Type>,sizeof(Type)> op; return op(x,y);}};
558: template<typename Type> struct AtomicLOR  {__device__ Type operator()(Type& x,Type y) const {AtomicLogical<Type,lor<Type> ,sizeof(Type)> op; return op(x,y);}};
559: template<typename Type> struct AtomicLXOR {__device__ Type operator()(Type& x,Type y) const {AtomicLogical<Type,lxor<Type>,sizeof(Type)> op; return op(x,y);}};

561: /*====================================================================================*/
562: /*  Wrapper functions of cuda kernels. Function pointers are stored in 'link'         */
563: /*====================================================================================*/
564: template<typename Type,PetscInt BS,PetscInt EQ>
565: static PetscErrorCode Pack(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,const void *data,void *buf)
566: {
567:   cudaError_t        cerr;
568:   PetscInt           nthreads=256;
569:   PetscInt           nblocks=(count+nthreads-1)/nthreads;
570:   const PetscInt     *iarray=opt ? opt->array : NULL;

573:   if (!count) return(0);
574:   if (!opt && !idx) { /* It is a 'CUDA data to nvshmem buf' memory copy */
575:     cerr = cudaMemcpyAsync(buf,(char*)data+start*link->unitbytes,count*link->unitbytes,cudaMemcpyDeviceToDevice,link->stream);CHKERRCUDA(cerr);
576:   } else {
577:     nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);
578:     d_Pack<Type,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,start,iarray,idx,(const Type*)data,(Type*)buf);
579:     cerr = cudaGetLastError();CHKERRCUDA(cerr);
580:   }
581:   return(0);
582: }

584: /* To specialize UnpackAndOp for the cudaMemcpyAsync() below. Usually if this is a contiguous memcpy, we use root/leafdirect and do
585:    not need UnpackAndOp. Only with nvshmem, we need this 'nvshmem buf to CUDA data' memory copy
586: */
587: template<typename Type,PetscInt BS,PetscInt EQ>
588: static PetscErrorCode Unpack(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *data,const void *buf)
589: {
590:   cudaError_t        cerr;
591:   PetscInt           nthreads=256;
592:   PetscInt           nblocks=(count+nthreads-1)/nthreads;
593:   const PetscInt     *iarray=opt ? opt->array : NULL;

596:   if (!count) return(0);
597:   if (!opt && !idx) { /* It is a 'nvshmem buf to CUDA data' memory copy */
598:     cerr = cudaMemcpyAsync((char*)data+start*link->unitbytes,buf,count*link->unitbytes,cudaMemcpyDeviceToDevice,link->stream);CHKERRCUDA(cerr);
599:   } else {
600:     nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);
601:     d_UnpackAndOp<Type,Insert<Type>,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,start,iarray,idx,(Type*)data,(const Type*)buf);
602:     cerr = cudaGetLastError();CHKERRCUDA(cerr);
603:   }
604:   return(0);
605: }

607: template<typename Type,class Op,PetscInt BS,PetscInt EQ>
608: static PetscErrorCode UnpackAndOp(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *data,const void *buf)
609: {
610:   cudaError_t        cerr;
611:   PetscInt           nthreads=256;
612:   PetscInt           nblocks=(count+nthreads-1)/nthreads;
613:   const PetscInt     *iarray=opt ? opt->array : NULL;

616:   if (!count) return(0);
617:   nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);
618:   d_UnpackAndOp<Type,Op,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,start,iarray,idx,(Type*)data,(const Type*)buf);
619:   cerr = cudaGetLastError();CHKERRCUDA(cerr);
620:   return(0);
621: }

623: template<typename Type,class Op,PetscInt BS,PetscInt EQ>
624: static PetscErrorCode FetchAndOp(PetscSFLink link,PetscInt count,PetscInt start,PetscSFPackOpt opt,const PetscInt *idx,void *data,void *buf)
625: {
626:   cudaError_t        cerr;
627:   PetscInt           nthreads=256;
628:   PetscInt           nblocks=(count+nthreads-1)/nthreads;
629:   const PetscInt     *iarray=opt ? opt->array : NULL;

632:   if (!count) return(0);
633:   nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);
634:   d_FetchAndOp<Type,Op,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,start,iarray,idx,(Type*)data,(Type*)buf);
635:   cerr = cudaGetLastError();CHKERRCUDA(cerr);
636:   return(0);
637: }

639: template<typename Type,class Op,PetscInt BS,PetscInt EQ>
640: static PetscErrorCode ScatterAndOp(PetscSFLink link,PetscInt count,PetscInt srcStart,PetscSFPackOpt srcOpt,const PetscInt *srcIdx,const void *src,PetscInt dstStart,PetscSFPackOpt dstOpt,const PetscInt *dstIdx,void *dst)
641: {
642:   cudaError_t        cerr;
643:   PetscInt           nthreads=256;
644:   PetscInt           nblocks=(count+nthreads-1)/nthreads;
645:   PetscInt           srcx=0,srcy=0,srcX=0,srcY=0,dstx=0,dsty=0,dstX=0,dstY=0;

648:   if (!count) return(0);
649:   nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);

651:   /* The 3D shape of source subdomain may be different than that of the destination, which makes it difficult to use CUDA 3D grid and block */
652:   if (srcOpt)       {srcx = srcOpt->dx[0]; srcy = srcOpt->dy[0]; srcX = srcOpt->X[0]; srcY = srcOpt->Y[0]; srcStart = srcOpt->start[0]; srcIdx = NULL;}
653:   else if (!srcIdx) {srcx = srcX = count; srcy = srcY = 1;}

655:   if (dstOpt)       {dstx = dstOpt->dx[0]; dsty = dstOpt->dy[0]; dstX = dstOpt->X[0]; dstY = dstOpt->Y[0]; dstStart = dstOpt->start[0]; dstIdx = NULL;}
656:   else if (!dstIdx) {dstx = dstX = count; dsty = dstY = 1;}

658:   d_ScatterAndOp<Type,Op,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,srcx,srcy,srcX,srcY,srcStart,srcIdx,(const Type*)src,dstx,dsty,dstX,dstY,dstStart,dstIdx,(Type*)dst);
659:   cerr = cudaGetLastError();CHKERRCUDA(cerr);
660:   return(0);
661: }

663: /* Specialization for Insert since we may use cudaMemcpyAsync */
664: template<typename Type,PetscInt BS,PetscInt EQ>
665: static PetscErrorCode ScatterAndInsert(PetscSFLink link,PetscInt count,PetscInt srcStart,PetscSFPackOpt srcOpt,const PetscInt *srcIdx,const void *src,PetscInt dstStart,PetscSFPackOpt dstOpt,const PetscInt *dstIdx,void *dst)
666: {
667:   PetscErrorCode    ierr;
668:   cudaError_t       cerr;

671:   if (!count) return(0);
672:   /*src and dst are contiguous */
673:   if ((!srcOpt && !srcIdx) && (!dstOpt && !dstIdx) && src != dst) {
674:     cerr = cudaMemcpyAsync((Type*)dst+dstStart*link->bs,(const Type*)src+srcStart*link->bs,count*link->unitbytes,cudaMemcpyDeviceToDevice,link->stream);CHKERRCUDA(cerr);
675:   } else {
676:     ScatterAndOp<Type,Insert<Type>,BS,EQ>(link,count,srcStart,srcOpt,srcIdx,src,dstStart,dstOpt,dstIdx,dst);
677:   }
678:   return(0);
679: }

681: template<typename Type,class Op,PetscInt BS,PetscInt EQ>
682: static PetscErrorCode FetchAndOpLocal(PetscSFLink link,PetscInt count,PetscInt rootstart,PetscSFPackOpt rootopt,const PetscInt *rootidx,void *rootdata,PetscInt leafstart,PetscSFPackOpt leafopt,const PetscInt *leafidx,const void *leafdata,void *leafupdate)
683: {
684:   cudaError_t       cerr;
685:   PetscInt          nthreads=256;
686:   PetscInt          nblocks=(count+nthreads-1)/nthreads;
687:   const PetscInt    *rarray = rootopt ? rootopt->array : NULL;
688:   const PetscInt    *larray = leafopt ? leafopt->array : NULL;

691:   if (!count) return(0);
692:   nblocks = PetscMin(nblocks,link->maxResidentThreadsPerGPU/nthreads);
693:   d_FetchAndOpLocal<Type,Op,BS,EQ><<<nblocks,nthreads,0,link->stream>>>(link->bs,count,rootstart,rarray,rootidx,(Type*)rootdata,leafstart,larray,leafidx,(const Type*)leafdata,(Type*)leafupdate);
694:   cerr = cudaGetLastError();CHKERRCUDA(cerr);
695:   return(0);
696: }

698: /*====================================================================================*/
699: /*  Init various types and instantiate pack/unpack function pointers                  */
700: /*====================================================================================*/
701: template<typename Type,PetscInt BS,PetscInt EQ>
702: static void PackInit_RealType(PetscSFLink link)
703: {
704:   /* Pack/unpack for remote communication */
705:   link->d_Pack              = Pack<Type,BS,EQ>;
706:   link->d_UnpackAndInsert   = Unpack<Type,BS,EQ>;
707:   link->d_UnpackAndAdd      = UnpackAndOp     <Type,Add<Type>         ,BS,EQ>;
708:   link->d_UnpackAndMult     = UnpackAndOp     <Type,Mult<Type>        ,BS,EQ>;
709:   link->d_UnpackAndMin      = UnpackAndOp     <Type,Min<Type>         ,BS,EQ>;
710:   link->d_UnpackAndMax      = UnpackAndOp     <Type,Max<Type>         ,BS,EQ>;
711:   link->d_FetchAndAdd       = FetchAndOp      <Type,Add<Type>         ,BS,EQ>;

713:   /* Scatter for local communication */
714:   link->d_ScatterAndInsert  = ScatterAndInsert<Type                   ,BS,EQ>; /* Has special optimizations */
715:   link->d_ScatterAndAdd     = ScatterAndOp    <Type,Add<Type>         ,BS,EQ>;
716:   link->d_ScatterAndMult    = ScatterAndOp    <Type,Mult<Type>        ,BS,EQ>;
717:   link->d_ScatterAndMin     = ScatterAndOp    <Type,Min<Type>         ,BS,EQ>;
718:   link->d_ScatterAndMax     = ScatterAndOp    <Type,Max<Type>         ,BS,EQ>;
719:   link->d_FetchAndAddLocal  = FetchAndOpLocal <Type,Add <Type>        ,BS,EQ>;

721:   /* Atomic versions when there are data-race possibilities */
722:   link->da_UnpackAndInsert  = UnpackAndOp     <Type,AtomicInsert<Type>,BS,EQ>;
723:   link->da_UnpackAndAdd     = UnpackAndOp     <Type,AtomicAdd<Type>   ,BS,EQ>;
724:   link->da_UnpackAndMult    = UnpackAndOp     <Type,AtomicMult<Type>  ,BS,EQ>;
725:   link->da_UnpackAndMin     = UnpackAndOp     <Type,AtomicMin<Type>   ,BS,EQ>;
726:   link->da_UnpackAndMax     = UnpackAndOp     <Type,AtomicMax<Type>   ,BS,EQ>;
727:   link->da_FetchAndAdd      = FetchAndOp      <Type,AtomicAdd<Type>   ,BS,EQ>;

729:   link->da_ScatterAndInsert = ScatterAndOp    <Type,AtomicInsert<Type>,BS,EQ>;
730:   link->da_ScatterAndAdd    = ScatterAndOp    <Type,AtomicAdd<Type>   ,BS,EQ>;
731:   link->da_ScatterAndMult   = ScatterAndOp    <Type,AtomicMult<Type>  ,BS,EQ>;
732:   link->da_ScatterAndMin    = ScatterAndOp    <Type,AtomicMin<Type>   ,BS,EQ>;
733:   link->da_ScatterAndMax    = ScatterAndOp    <Type,AtomicMax<Type>   ,BS,EQ>;
734:   link->da_FetchAndAddLocal = FetchAndOpLocal <Type,AtomicAdd<Type>   ,BS,EQ>;
735: }

737: /* Have this templated class to specialize for char integers */
738: template<typename Type,PetscInt BS,PetscInt EQ,PetscInt size/*sizeof(Type)*/>
739: struct PackInit_IntegerType_Atomic {
740:   static void Init(PetscSFLink link) {
741:     link->da_UnpackAndInsert  = UnpackAndOp<Type,AtomicInsert<Type>,BS,EQ>;
742:     link->da_UnpackAndAdd     = UnpackAndOp<Type,AtomicAdd<Type>   ,BS,EQ>;
743:     link->da_UnpackAndMult    = UnpackAndOp<Type,AtomicMult<Type>  ,BS,EQ>;
744:     link->da_UnpackAndMin     = UnpackAndOp<Type,AtomicMin<Type>   ,BS,EQ>;
745:     link->da_UnpackAndMax     = UnpackAndOp<Type,AtomicMax<Type>   ,BS,EQ>;
746:     link->da_UnpackAndLAND    = UnpackAndOp<Type,AtomicLAND<Type>  ,BS,EQ>;
747:     link->da_UnpackAndLOR     = UnpackAndOp<Type,AtomicLOR<Type>   ,BS,EQ>;
748:     link->da_UnpackAndLXOR    = UnpackAndOp<Type,AtomicLXOR<Type>  ,BS,EQ>;
749:     link->da_UnpackAndBAND    = UnpackAndOp<Type,AtomicBAND<Type>  ,BS,EQ>;
750:     link->da_UnpackAndBOR     = UnpackAndOp<Type,AtomicBOR<Type>   ,BS,EQ>;
751:     link->da_UnpackAndBXOR    = UnpackAndOp<Type,AtomicBXOR<Type>  ,BS,EQ>;
752:     link->da_FetchAndAdd      = FetchAndOp <Type,AtomicAdd<Type>   ,BS,EQ>;

754:     link->da_ScatterAndInsert = ScatterAndOp<Type,AtomicInsert<Type>,BS,EQ>;
755:     link->da_ScatterAndAdd    = ScatterAndOp<Type,AtomicAdd<Type>   ,BS,EQ>;
756:     link->da_ScatterAndMult   = ScatterAndOp<Type,AtomicMult<Type>  ,BS,EQ>;
757:     link->da_ScatterAndMin    = ScatterAndOp<Type,AtomicMin<Type>   ,BS,EQ>;
758:     link->da_ScatterAndMax    = ScatterAndOp<Type,AtomicMax<Type>   ,BS,EQ>;
759:     link->da_ScatterAndLAND   = ScatterAndOp<Type,AtomicLAND<Type>  ,BS,EQ>;
760:     link->da_ScatterAndLOR    = ScatterAndOp<Type,AtomicLOR<Type>   ,BS,EQ>;
761:     link->da_ScatterAndLXOR   = ScatterAndOp<Type,AtomicLXOR<Type>  ,BS,EQ>;
762:     link->da_ScatterAndBAND   = ScatterAndOp<Type,AtomicBAND<Type>  ,BS,EQ>;
763:     link->da_ScatterAndBOR    = ScatterAndOp<Type,AtomicBOR<Type>   ,BS,EQ>;
764:     link->da_ScatterAndBXOR   = ScatterAndOp<Type,AtomicBXOR<Type>  ,BS,EQ>;
765:     link->da_FetchAndAddLocal = FetchAndOpLocal<Type,AtomicAdd<Type>,BS,EQ>;
766:   }
767: };

769: /* CUDA does not support atomics on chars. It is TBD in PETSc. */
770: template<typename Type,PetscInt BS,PetscInt EQ>
771: struct PackInit_IntegerType_Atomic<Type,BS,EQ,1> {
772:   static void Init(PetscSFLink link) {/* Nothing to leave function pointers NULL */}
773: };

775: template<typename Type,PetscInt BS,PetscInt EQ>
776: static void PackInit_IntegerType(PetscSFLink link)
777: {
778:   link->d_Pack            = Pack<Type,BS,EQ>;
779:   link->d_UnpackAndInsert = Unpack<Type,BS,EQ>;
780:   link->d_UnpackAndAdd    = UnpackAndOp<Type,Add<Type>   ,BS,EQ>;
781:   link->d_UnpackAndMult   = UnpackAndOp<Type,Mult<Type>  ,BS,EQ>;
782:   link->d_UnpackAndMin    = UnpackAndOp<Type,Min<Type>   ,BS,EQ>;
783:   link->d_UnpackAndMax    = UnpackAndOp<Type,Max<Type>   ,BS,EQ>;
784:   link->d_UnpackAndLAND   = UnpackAndOp<Type,LAND<Type>  ,BS,EQ>;
785:   link->d_UnpackAndLOR    = UnpackAndOp<Type,LOR<Type>   ,BS,EQ>;
786:   link->d_UnpackAndLXOR   = UnpackAndOp<Type,LXOR<Type>  ,BS,EQ>;
787:   link->d_UnpackAndBAND   = UnpackAndOp<Type,BAND<Type>  ,BS,EQ>;
788:   link->d_UnpackAndBOR    = UnpackAndOp<Type,BOR<Type>   ,BS,EQ>;
789:   link->d_UnpackAndBXOR   = UnpackAndOp<Type,BXOR<Type>  ,BS,EQ>;
790:   link->d_FetchAndAdd     = FetchAndOp <Type,Add<Type>   ,BS,EQ>;

792:   link->d_ScatterAndInsert = ScatterAndInsert<Type,BS,EQ>;
793:   link->d_ScatterAndAdd    = ScatterAndOp<Type,Add<Type>   ,BS,EQ>;
794:   link->d_ScatterAndMult   = ScatterAndOp<Type,Mult<Type>  ,BS,EQ>;
795:   link->d_ScatterAndMin    = ScatterAndOp<Type,Min<Type>   ,BS,EQ>;
796:   link->d_ScatterAndMax    = ScatterAndOp<Type,Max<Type>   ,BS,EQ>;
797:   link->d_ScatterAndLAND   = ScatterAndOp<Type,LAND<Type>  ,BS,EQ>;
798:   link->d_ScatterAndLOR    = ScatterAndOp<Type,LOR<Type>   ,BS,EQ>;
799:   link->d_ScatterAndLXOR   = ScatterAndOp<Type,LXOR<Type>  ,BS,EQ>;
800:   link->d_ScatterAndBAND   = ScatterAndOp<Type,BAND<Type>  ,BS,EQ>;
801:   link->d_ScatterAndBOR    = ScatterAndOp<Type,BOR<Type>   ,BS,EQ>;
802:   link->d_ScatterAndBXOR   = ScatterAndOp<Type,BXOR<Type>  ,BS,EQ>;
803:   link->d_FetchAndAddLocal = FetchAndOpLocal<Type,Add<Type>,BS,EQ>;
804:   PackInit_IntegerType_Atomic<Type,BS,EQ,sizeof(Type)>::Init(link);
805: }

807: #if defined(PETSC_HAVE_COMPLEX)
808: template<typename Type,PetscInt BS,PetscInt EQ>
809: static void PackInit_ComplexType(PetscSFLink link)
810: {
811:   link->d_Pack             = Pack<Type,BS,EQ>;
812:   link->d_UnpackAndInsert  = Unpack<Type,BS,EQ>;
813:   link->d_UnpackAndAdd     = UnpackAndOp<Type,Add<Type>   ,BS,EQ>;
814:   link->d_UnpackAndMult    = UnpackAndOp<Type,Mult<Type>  ,BS,EQ>;
815:   link->d_FetchAndAdd      = FetchAndOp <Type,Add<Type>   ,BS,EQ>;

817:   link->d_ScatterAndInsert = ScatterAndInsert<Type,BS,EQ>;
818:   link->d_ScatterAndAdd    = ScatterAndOp<Type,Add<Type>   ,BS,EQ>;
819:   link->d_ScatterAndMult   = ScatterAndOp<Type,Mult<Type>  ,BS,EQ>;
820:   link->d_FetchAndAddLocal = FetchAndOpLocal<Type,Add<Type>,BS,EQ>;

822:   link->da_UnpackAndInsert = UnpackAndOp<Type,AtomicInsert<Type>,BS,EQ>;
823:   link->da_UnpackAndAdd    = UnpackAndOp<Type,AtomicAdd<Type>,BS,EQ>;
824:   link->da_UnpackAndMult   = NULL; /* Not implemented yet */
825:   link->da_FetchAndAdd     = NULL; /* Return value of atomicAdd on complex is not atomic */

827:   link->da_ScatterAndInsert = ScatterAndOp<Type,AtomicInsert<Type>,BS,EQ>;
828:   link->da_ScatterAndAdd    = ScatterAndOp<Type,AtomicAdd<Type>,BS,EQ>;
829: }
830: #endif

832: typedef signed char                      SignedChar;
833: typedef unsigned char                    UnsignedChar;
834: typedef struct {int a;      int b;     } PairInt;
835: typedef struct {PetscInt a; PetscInt b;} PairPetscInt;

837: template<typename Type>
838: static void PackInit_PairType(PetscSFLink link)
839: {
840:   link->d_Pack            = Pack<Type,1,1>;
841:   link->d_UnpackAndInsert = Unpack<Type,1,1>;
842:   link->d_UnpackAndMaxloc = UnpackAndOp<Type,Maxloc<Type>,1,1>;
843:   link->d_UnpackAndMinloc = UnpackAndOp<Type,Minloc<Type>,1,1>;

845:   link->d_ScatterAndInsert = ScatterAndOp<Type,Insert<Type>,1,1>;
846:   link->d_ScatterAndMaxloc = ScatterAndOp<Type,Maxloc<Type>,1,1>;
847:   link->d_ScatterAndMinloc = ScatterAndOp<Type,Minloc<Type>,1,1>;
848:   /* Atomics for pair types are not implemented yet */
849: }

851: template<typename Type,PetscInt BS,PetscInt EQ>
852: static void PackInit_DumbType(PetscSFLink link)
853: {
854:   link->d_Pack             = Pack<Type,BS,EQ>;
855:   link->d_UnpackAndInsert  = Unpack<Type,BS,EQ>;
856:   link->d_ScatterAndInsert = ScatterAndInsert<Type,BS,EQ>;
857:   /* Atomics for dumb types are not implemented yet */
858: }

860: /* Some device-specific utilities */
861: static PetscErrorCode PetscSFLinkSyncDevice_CUDA(PetscSFLink link)
862: {
863:   cudaError_t cerr;
865:   cerr = cudaDeviceSynchronize();CHKERRCUDA(cerr);
866:   return(0);
867: }

869: static PetscErrorCode PetscSFLinkSyncStream_CUDA(PetscSFLink link)
870: {
871:   cudaError_t cerr;
873:   cerr = cudaStreamSynchronize(link->stream);CHKERRCUDA(cerr);
874:   return(0);
875: }

877: static PetscErrorCode PetscSFLinkMemcpy_CUDA(PetscSFLink link,PetscMemType dstmtype,void* dst,PetscMemType srcmtype,const void*src,size_t n)
878: {
880:   enum cudaMemcpyKind kinds[2][2] = {{cudaMemcpyHostToHost,cudaMemcpyHostToDevice},{cudaMemcpyDeviceToHost,cudaMemcpyDeviceToDevice}};

882:   if (n) {
883:     if (PetscMemTypeHost(dstmtype) && PetscMemTypeHost(srcmtype)) { /* Separate HostToHost so that pure-cpu code won't call cuda runtime */
884:       PetscErrorCode PetscMemcpy(dst,src,n);
885:     } else {
886:       int stype = PetscMemTypeDevice(srcmtype) ? 1 : 0;
887:       int dtype = PetscMemTypeDevice(dstmtype) ? 1 : 0;
888:       cudaError_t cerr = cudaMemcpyAsync(dst,src,n,kinds[stype][dtype],link->stream);CHKERRCUDA(cerr);
889:     }
890:   }
891:   return(0);
892: }

894: PetscErrorCode PetscSFMalloc_CUDA(PetscMemType mtype,size_t size,void** ptr)
895: {
897:   if (PetscMemTypeHost(mtype)) {PetscErrorCode PetscMalloc(size,ptr);}
898:   else if (PetscMemTypeDevice(mtype)) {
899:     if (!PetscCUDAInitialized) {PetscErrorCode PetscCUDAInitializeCheck();}
900:     cudaError_t err = cudaMalloc(ptr,size);CHKERRCUDA(err);
901:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Wrong PetscMemType %d", (int)mtype);
902:   return(0);
903: }

905: PetscErrorCode PetscSFFree_CUDA(PetscMemType mtype,void* ptr)
906: {
908:   if (PetscMemTypeHost(mtype)) {PetscErrorCode PetscFree(ptr);}
909:   else if (PetscMemTypeDevice(mtype)) {cudaError_t err = cudaFree(ptr);CHKERRCUDA(err);}
910:   else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Wrong PetscMemType %d",(int)mtype);
911:   return(0);
912: }

914: /* Destructor when the link uses MPI for communication on CUDA device */
915: static PetscErrorCode PetscSFLinkDestroy_MPI_CUDA(PetscSF sf,PetscSFLink link)
916: {
917:   cudaError_t    cerr;

920:   for (int i=PETSCSF_LOCAL; i<=PETSCSF_REMOTE; i++) {
921:     cerr = cudaFree(link->rootbuf_alloc[i][PETSC_MEMTYPE_DEVICE]);CHKERRCUDA(cerr);
922:     cerr = cudaFree(link->leafbuf_alloc[i][PETSC_MEMTYPE_DEVICE]);CHKERRCUDA(cerr);
923:   }
924:   return(0);
925: }

927: /* Some fields of link are initialized by PetscSFPackSetUp_Host. This routine only does what needed on device */
928: PetscErrorCode PetscSFLinkSetUp_CUDA(PetscSF sf,PetscSFLink link,MPI_Datatype unit)
929: {
931:   cudaError_t    cerr;
932:   PetscInt       nSignedChar=0,nUnsignedChar=0,nInt=0,nPetscInt=0,nPetscReal=0;
933:   PetscBool      is2Int,is2PetscInt;
934: #if defined(PETSC_HAVE_COMPLEX)
935:   PetscInt       nPetscComplex=0;
936: #endif

939:   if (link->deviceinited) return(0);
940:   MPIPetsc_Type_compare_contig(unit,MPI_SIGNED_CHAR,  &nSignedChar);
941:   MPIPetsc_Type_compare_contig(unit,MPI_UNSIGNED_CHAR,&nUnsignedChar);
942:   /* MPI_CHAR is treated below as a dumb type that does not support reduction according to MPI standard */
943:   MPIPetsc_Type_compare_contig(unit,MPI_INT,  &nInt);
944:   MPIPetsc_Type_compare_contig(unit,MPIU_INT, &nPetscInt);
945:   MPIPetsc_Type_compare_contig(unit,MPIU_REAL,&nPetscReal);
946: #if defined(PETSC_HAVE_COMPLEX)
947:   MPIPetsc_Type_compare_contig(unit,MPIU_COMPLEX,&nPetscComplex);
948: #endif
949:   MPIPetsc_Type_compare(unit,MPI_2INT,&is2Int);
950:   MPIPetsc_Type_compare(unit,MPIU_2INT,&is2PetscInt);

952:   if (is2Int) {
953:     PackInit_PairType<PairInt>(link);
954:   } else if (is2PetscInt) { /* TODO: when is2PetscInt and nPetscInt=2, we don't know which path to take. The two paths support different ops. */
955:     PackInit_PairType<PairPetscInt>(link);
956:   } else if (nPetscReal) {
957:     if      (nPetscReal == 8) PackInit_RealType<PetscReal,8,1>(link); else if (nPetscReal%8 == 0) PackInit_RealType<PetscReal,8,0>(link);
958:     else if (nPetscReal == 4) PackInit_RealType<PetscReal,4,1>(link); else if (nPetscReal%4 == 0) PackInit_RealType<PetscReal,4,0>(link);
959:     else if (nPetscReal == 2) PackInit_RealType<PetscReal,2,1>(link); else if (nPetscReal%2 == 0) PackInit_RealType<PetscReal,2,0>(link);
960:     else if (nPetscReal == 1) PackInit_RealType<PetscReal,1,1>(link); else if (nPetscReal%1 == 0) PackInit_RealType<PetscReal,1,0>(link);
961:   } else if (nPetscInt && sizeof(PetscInt) == sizeof(llint)) {
962:     if      (nPetscInt == 8) PackInit_IntegerType<llint,8,1>(link); else if (nPetscInt%8 == 0) PackInit_IntegerType<llint,8,0>(link);
963:     else if (nPetscInt == 4) PackInit_IntegerType<llint,4,1>(link); else if (nPetscInt%4 == 0) PackInit_IntegerType<llint,4,0>(link);
964:     else if (nPetscInt == 2) PackInit_IntegerType<llint,2,1>(link); else if (nPetscInt%2 == 0) PackInit_IntegerType<llint,2,0>(link);
965:     else if (nPetscInt == 1) PackInit_IntegerType<llint,1,1>(link); else if (nPetscInt%1 == 0) PackInit_IntegerType<llint,1,0>(link);
966:   } else if (nInt) {
967:     if      (nInt == 8) PackInit_IntegerType<int,8,1>(link); else if (nInt%8 == 0) PackInit_IntegerType<int,8,0>(link);
968:     else if (nInt == 4) PackInit_IntegerType<int,4,1>(link); else if (nInt%4 == 0) PackInit_IntegerType<int,4,0>(link);
969:     else if (nInt == 2) PackInit_IntegerType<int,2,1>(link); else if (nInt%2 == 0) PackInit_IntegerType<int,2,0>(link);
970:     else if (nInt == 1) PackInit_IntegerType<int,1,1>(link); else if (nInt%1 == 0) PackInit_IntegerType<int,1,0>(link);
971:   } else if (nSignedChar) {
972:     if      (nSignedChar == 8) PackInit_IntegerType<SignedChar,8,1>(link); else if (nSignedChar%8 == 0) PackInit_IntegerType<SignedChar,8,0>(link);
973:     else if (nSignedChar == 4) PackInit_IntegerType<SignedChar,4,1>(link); else if (nSignedChar%4 == 0) PackInit_IntegerType<SignedChar,4,0>(link);
974:     else if (nSignedChar == 2) PackInit_IntegerType<SignedChar,2,1>(link); else if (nSignedChar%2 == 0) PackInit_IntegerType<SignedChar,2,0>(link);
975:     else if (nSignedChar == 1) PackInit_IntegerType<SignedChar,1,1>(link); else if (nSignedChar%1 == 0) PackInit_IntegerType<SignedChar,1,0>(link);
976:   }  else if (nUnsignedChar) {
977:     if      (nUnsignedChar == 8) PackInit_IntegerType<UnsignedChar,8,1>(link); else if (nUnsignedChar%8 == 0) PackInit_IntegerType<UnsignedChar,8,0>(link);
978:     else if (nUnsignedChar == 4) PackInit_IntegerType<UnsignedChar,4,1>(link); else if (nUnsignedChar%4 == 0) PackInit_IntegerType<UnsignedChar,4,0>(link);
979:     else if (nUnsignedChar == 2) PackInit_IntegerType<UnsignedChar,2,1>(link); else if (nUnsignedChar%2 == 0) PackInit_IntegerType<UnsignedChar,2,0>(link);
980:     else if (nUnsignedChar == 1) PackInit_IntegerType<UnsignedChar,1,1>(link); else if (nUnsignedChar%1 == 0) PackInit_IntegerType<UnsignedChar,1,0>(link);
981: #if defined(PETSC_HAVE_COMPLEX)
982:   } else if (nPetscComplex) {
983:     if      (nPetscComplex == 8) PackInit_ComplexType<PetscComplex,8,1>(link); else if (nPetscComplex%8 == 0) PackInit_ComplexType<PetscComplex,8,0>(link);
984:     else if (nPetscComplex == 4) PackInit_ComplexType<PetscComplex,4,1>(link); else if (nPetscComplex%4 == 0) PackInit_ComplexType<PetscComplex,4,0>(link);
985:     else if (nPetscComplex == 2) PackInit_ComplexType<PetscComplex,2,1>(link); else if (nPetscComplex%2 == 0) PackInit_ComplexType<PetscComplex,2,0>(link);
986:     else if (nPetscComplex == 1) PackInit_ComplexType<PetscComplex,1,1>(link); else if (nPetscComplex%1 == 0) PackInit_ComplexType<PetscComplex,1,0>(link);
987: #endif
988:   } else {
989:     MPI_Aint lb,nbyte;
990:     MPI_Type_get_extent(unit,&lb,&nbyte);
991:     if (lb != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
992:     if (nbyte % sizeof(int)) { /* If the type size is not multiple of int */
993:       if      (nbyte == 4) PackInit_DumbType<char,4,1>(link); else if (nbyte%4 == 0) PackInit_DumbType<char,4,0>(link);
994:       else if (nbyte == 2) PackInit_DumbType<char,2,1>(link); else if (nbyte%2 == 0) PackInit_DumbType<char,2,0>(link);
995:       else if (nbyte == 1) PackInit_DumbType<char,1,1>(link); else if (nbyte%1 == 0) PackInit_DumbType<char,1,0>(link);
996:     } else {
997:       nInt = nbyte / sizeof(int);
998:       if      (nInt == 8) PackInit_DumbType<int,8,1>(link); else if (nInt%8 == 0) PackInit_DumbType<int,8,0>(link);
999:       else if (nInt == 4) PackInit_DumbType<int,4,1>(link); else if (nInt%4 == 0) PackInit_DumbType<int,4,0>(link);
1000:       else if (nInt == 2) PackInit_DumbType<int,2,1>(link); else if (nInt%2 == 0) PackInit_DumbType<int,2,0>(link);
1001:       else if (nInt == 1) PackInit_DumbType<int,1,1>(link); else if (nInt%1 == 0) PackInit_DumbType<int,1,0>(link);
1002:     }
1003:   }

1005:   if (!sf->maxResidentThreadsPerGPU) { /* Not initialized */
1006:     int                   device;
1007:     struct cudaDeviceProp props;
1008:     cerr = cudaGetDevice(&device);CHKERRCUDA(cerr);
1009:     cerr = cudaGetDeviceProperties(&props,device);CHKERRCUDA(cerr);
1010:     sf->maxResidentThreadsPerGPU = props.maxThreadsPerMultiProcessor*props.multiProcessorCount;
1011:   }
1012:   link->maxResidentThreadsPerGPU = sf->maxResidentThreadsPerGPU;

1014:   link->stream             = PetscDefaultCudaStream;
1015:   link->Destroy            = PetscSFLinkDestroy_MPI_CUDA;
1016:   link->SyncDevice         = PetscSFLinkSyncDevice_CUDA;
1017:   link->SyncStream         = PetscSFLinkSyncStream_CUDA;
1018:   link->Memcpy             = PetscSFLinkMemcpy_CUDA;
1019:   link->deviceinited       = PETSC_TRUE;
1020:   return(0);
1021: }