1: #pragma once

  3: #include <petscdevicetypes.h>
  4: #include <petscviewertypes.h>

  6: #if PETSC_CPP_VERSION >= 11 // C++11
  7:   #define PETSC_DEVICE_ALIGNOF(...) alignof(decltype(__VA_ARGS__))
  8: #elif PETSC_C_VERSION >= 11 // C11
  9:   #ifdef __GNUC__
 10:     #define PETSC_DEVICE_ALIGNOF(...) _Alignof(__typeof__(__VA_ARGS__))
 11:   #else
 12:     #include <stddef.h> // max_align_t
 13:     // Note we cannot just do _Alignof(expression) since clang warns that "'_Alignof' applied to an
 14:     // expression is a GNU extension", so we just default to max_align_t which is ultra safe
 15:     #define PETSC_DEVICE_ALIGNOF(...) _Alignof(max_align_t)
 16:   #endif // __GNUC__
 17: #else
 18:   #define PETSC_DEVICE_ALIGNOF(...) PETSC_MEMALIGN
 19: #endif

 21: /* MANSEC = Sys */
 22: /* SUBMANSEC = Device */

 24: // REVIEW ME: this should probably go somewhere better, configure-time?
 25: #define PETSC_HAVE_HOST 1

 27: /* logging support */
 28: PETSC_EXTERN PetscClassId PETSC_DEVICE_CLASSID;
 29: PETSC_EXTERN PetscClassId PETSC_DEVICE_CONTEXT_CLASSID;

 31: PETSC_EXTERN PetscErrorCode PetscDeviceInitializePackage(void);
 32: PETSC_EXTERN PetscErrorCode PetscDeviceFinalizePackage(void);
 33: PETSC_EXTERN PetscErrorCode PetscGetMemType(const void *, PetscMemType *);

 35: /* PetscDevice */
 36: PETSC_EXTERN PetscErrorCode  PetscDeviceCreate(PetscDeviceType, PetscInt, PetscDevice *);
 37: PETSC_EXTERN PetscErrorCode  PetscDeviceDestroy(PetscDevice *);
 38: PETSC_EXTERN PetscErrorCode  PetscDeviceConfigure(PetscDevice);
 39: PETSC_EXTERN PetscErrorCode  PetscDeviceView(PetscDevice, PetscViewer);
 40: PETSC_EXTERN PetscErrorCode  PetscDeviceGetType(PetscDevice, PetscDeviceType *);
 41: PETSC_EXTERN PetscErrorCode  PetscDeviceGetDeviceId(PetscDevice, PetscInt *);
 42: PETSC_EXTERN PetscDeviceType PETSC_DEVICE_DEFAULT(void);
 43: PETSC_EXTERN PetscErrorCode  PetscDeviceSetDefaultDeviceType(PetscDeviceType);
 44: PETSC_EXTERN PetscErrorCode  PetscDeviceInitialize(PetscDeviceType);
 45: PETSC_EXTERN PetscBool       PetscDeviceInitialized(PetscDeviceType);

 47: /* PetscDeviceContext */
 48: PETSC_EXTERN PetscErrorCode PetscDeviceContextCreate(PetscDeviceContext *);
 49: PETSC_EXTERN PetscErrorCode PetscDeviceContextDestroy(PetscDeviceContext *);
 50: PETSC_EXTERN PetscErrorCode PetscDeviceContextSetStreamType(PetscDeviceContext, PetscStreamType);
 51: PETSC_EXTERN PetscErrorCode PetscDeviceContextGetStreamType(PetscDeviceContext, PetscStreamType *);
 52: PETSC_EXTERN PetscErrorCode PetscDeviceContextSetDevice(PetscDeviceContext, PetscDevice);
 53: PETSC_EXTERN PetscErrorCode PetscDeviceContextGetDevice(PetscDeviceContext, PetscDevice *);
 54: PETSC_EXTERN PetscErrorCode PetscDeviceContextGetDeviceType(PetscDeviceContext, PetscDeviceType *);
 55: PETSC_EXTERN PetscErrorCode PetscDeviceContextSetUp(PetscDeviceContext);
 56: PETSC_EXTERN PetscErrorCode PetscDeviceContextDuplicate(PetscDeviceContext, PetscDeviceContext *);
 57: PETSC_EXTERN PetscErrorCode PetscDeviceContextQueryIdle(PetscDeviceContext, PetscBool *);
 58: PETSC_EXTERN PetscErrorCode PetscDeviceContextWaitForContext(PetscDeviceContext, PetscDeviceContext);
 59: PETSC_EXTERN PetscErrorCode PetscDeviceContextForkWithStreamType(PetscDeviceContext, PetscStreamType, PetscInt, PetscDeviceContext **);
 60: PETSC_EXTERN PetscErrorCode PetscDeviceContextFork(PetscDeviceContext, PetscInt, PetscDeviceContext **);
 61: PETSC_EXTERN PetscErrorCode PetscDeviceContextJoin(PetscDeviceContext, PetscInt, PetscDeviceContextJoinMode, PetscDeviceContext **);
 62: PETSC_EXTERN PetscErrorCode PetscDeviceContextSynchronize(PetscDeviceContext);
 63: PETSC_EXTERN PetscErrorCode PetscDeviceContextSetFromOptions(MPI_Comm, PetscDeviceContext);
 64: PETSC_EXTERN PetscErrorCode PetscDeviceContextView(PetscDeviceContext, PetscViewer);
 65: PETSC_EXTERN PetscErrorCode PetscDeviceContextViewFromOptions(PetscDeviceContext, PetscObject, const char[]);
 66: PETSC_EXTERN PetscErrorCode PetscDeviceContextGetCurrentContext(PetscDeviceContext *);
 67: PETSC_EXTERN PetscErrorCode PetscDeviceContextSetCurrentContext(PetscDeviceContext);
 68: PETSC_EXTERN PetscErrorCode PetscDeviceContextGetStreamHandle(PetscDeviceContext, void **);

 70: /* memory */
 71: PETSC_EXTERN PetscErrorCode PetscDeviceAllocate_Private(PetscDeviceContext, PetscBool, PetscMemType, size_t, size_t, void **PETSC_RESTRICT);
 72: PETSC_EXTERN PetscErrorCode PetscDeviceDeallocate_Private(PetscDeviceContext, void *PETSC_RESTRICT);
 73: PETSC_EXTERN PetscErrorCode PetscDeviceMemcpy(PetscDeviceContext, void *PETSC_RESTRICT, const void *PETSC_RESTRICT, size_t);
 74: PETSC_EXTERN PetscErrorCode PetscDeviceMemset(PetscDeviceContext, void *PETSC_RESTRICT, PetscInt, size_t);

 76: /*MC
 77:   PetscDeviceMalloc - Allocate device-aware memory

 79:   Synopsis:
 80: #include <petscdevice.h>
 81:   PetscErrorCode PetscDeviceMalloc(PetscDeviceContext dctx, PetscMemType mtype, size_t n, Type **ptr)

 83:   Not Collective, Asynchronous, Auto-dependency aware

 85:   Input Parameters:
 86: + dctx  - The `PetscDeviceContext` used to allocate the memory
 87: . mtype - The type of memory to allocate
 88: - n     - The amount (in elements) to allocate

 90:   Output Parameter:
 91: . ptr - The pointer to store the result in

 93:   Level: beginner

 95:   Notes:
 96:   Memory allocated with this function must be freed with `PetscDeviceFree()`.

 98:   If `n` is zero, then `ptr` is set to `PETSC_NULLPTR`.

100:   This routine falls back to using `PetscMalloc1()` if PETSc was not configured with device
101:   support. The user should note that `mtype` is ignored in this case, as `PetscMalloc1()`
102:   allocates only host memory.

104:   This routine uses the `sizeof()` of the memory type requested to determine the total memory
105:   to be allocated, therefore you should not multiply the number of elements requested by the
106:   `sizeof()` the type\:

108: .vb
109:   PetscInt *arr;

111:   // correct
112:   PetscDeviceMalloc(dctx,PETSC_MEMTYPE_DEVICE,n,&arr);

114:   // incorrect
115:   PetscDeviceMalloc(dctx,PETSC_MEMTYPE_DEVICE,n*sizeof(*arr),&arr);
116: .ve

118:   Note result stored `ptr` is immediately valid and the user may freely inspect or manipulate
119:   its value on function return, i.e.\:

121: .vb
122:   PetscInt *ptr;

124:   PetscDeviceMalloc(dctx, PETSC_MEMTYPE_DEVICE, 20, &ptr);

126:   PetscInt *sub_ptr = ptr + 10; // OK, no need to synchronize

128:   ptr[0] = 10; // ERROR, directly accessing contents of ptr is undefined until synchronization
129: .ve

131:   DAG representation:
132: .vb
133:   time ->

135:   -> dctx - |= CALL =| -\- dctx -->
136:                          \- ptr ->
137: .ve

139: .N ASYNC_API

141: .seealso: `PetscDeviceFree()`, `PetscDeviceCalloc()`, `PetscDeviceArrayCopy()`,
142: `PetscDeviceArrayZero()`
143: M*/
144: #define PetscDeviceMalloc(dctx, mtype, n, ptr) PetscDeviceAllocate_Private((dctx), PETSC_FALSE, (mtype), (size_t)(n) * sizeof(**(ptr)), PETSC_DEVICE_ALIGNOF(**(ptr)), (void **)(ptr))

146: /*MC
147:   PetscDeviceCalloc - Allocate zeroed device-aware memory

149:   Synopsis:
150: #include <petscdevice.h>
151:   PetscErrorCode PetscDeviceCalloc(PetscDeviceContext dctx, PetscMemType mtype, size_t n, Type **ptr)

153:   Not Collective, Asynchronous, Auto-dependency aware

155:   Input Parameters:
156: + dctx  - The `PetscDeviceContext` used to allocate the memory
157: . mtype - The type of memory to allocate
158: - n     - The amount (in elements) to allocate

160:   Output Parameter:
161: . ptr - The pointer to store the result in

163:   Level: beginner

165:   Notes:
166:   Has identical usage to `PetscDeviceMalloc()` except that the memory is zeroed before it is
167:   returned. See `PetscDeviceMalloc()` for further discussion.

169:   This routine falls back to using `PetscCalloc1()` if PETSc was not configured with device
170:   support. The user should note that `mtype` is ignored in this case, as `PetscCalloc1()`
171:   allocates only host memory.

173: .N ASYNC_API

175: .seealso: `PetscDeviceFree()`, `PetscDeviceMalloc()`, `PetscDeviceArrayCopy()`,
176: `PetscDeviceArrayZero()`
177: M*/
178: #define PetscDeviceCalloc(dctx, mtype, n, ptr) PetscDeviceAllocate_Private((dctx), PETSC_TRUE, (mtype), (size_t)(n) * sizeof(**(ptr)), PETSC_DEVICE_ALIGNOF(**(ptr)), (void **)(ptr))

180: /*MC
181:   PetscDeviceFree - Free device-aware memory obtained with  `PetscDeviceMalloc()` or `PetscDeviceCalloc()`

183:   Synopsis:
184: #include <petscdevice.h>
185:   PetscErrorCode PetscDeviceFree(PetscDeviceContext dctx, void *ptr)

187:   Not Collective, Asynchronous, Auto-dependency aware

189:   Input Parameters:
190: + dctx - The `PetscDeviceContext` used to free the memory
191: - ptr  - The pointer to free, may be `NULL`

193:   Level: beginner

195:   Notes:
196:   `ptr` is set to `PETSC_NULLPTR` on successful deallocation.

198:   `ptr` must have been allocated using `PetscDeviceMalloc()`, `PetscDeviceCalloc()` not `PetscMalloc()` or related routines

200:   This routine falls back to using `PetscFree()` if PETSc was not configured with device
201:   support. The user should note that `PetscFree()` frees only host memory.

203:   DAG representation:
204: .vb
205:   time ->

207:   -> dctx -/- |= CALL =| - dctx ->
208:   -> ptr -/
209: .ve

211: .N ASYNC_API

213: .seealso: `PetscDeviceMalloc()`, `PetscDeviceCalloc()`
214: M*/
215: #define PetscDeviceFree(dctx, ptr) ((PetscErrorCode)(PetscDeviceDeallocate_Private((dctx), (ptr)) || ((ptr) = PETSC_NULLPTR, PETSC_SUCCESS)))

217: /*MC
218:   PetscDeviceArrayCopy - Copy memory in a device-aware manner

220:   Synopsis:
221: #include <petscdevice.h>
222:   PetscErrorCode PetscDeviceArrayCopy(PetscDeviceContext dctx, void *dest, const void *src, size_t n)

224:   Not Collective, Asynchronous, Auto-dependency aware

226:   Input Parameters:
227: + dctx - The `PetscDeviceContext` used to copy the memory
228: . dest - The pointer to copy to
229: . src  - The pointer to copy from
230: - n    - The amount (in elements) to copy

232:   Notes:
233:   Both `dest` and `src` must have been allocated using `PetscDeviceMalloc()` or
234:   `PetscDeviceCalloc()`.

236:   This uses the `sizeof()` of the `src` memory type requested to determine the total memory to
237:   be copied, therefore you should not multiply the number of elements by the `sizeof()` the
238:   type\:

240: .vb
241:   PetscInt *to,*from;

243:   // correct
244:   PetscDeviceArrayCopy(dctx,to,from,n);

246:   // incorrect
247:   PetscDeviceArrayCopy(dctx,to,from,n*sizeof(*from));
248: .ve

250:   See `PetscDeviceMemcpy()` for further discussion.

252:   Level: beginner

254: .N ASYNC_API

256: .seealso: `PetscDeviceMalloc()`, `PetscDeviceCalloc()`, `PetscDeviceFree()`,
257: `PetscDeviceArrayZero()`, `PetscDeviceMemcpy()`
258: M*/
259: #define PetscDeviceArrayCopy(dctx, dest, src, n) PetscDeviceMemcpy((dctx), (dest), (src), (size_t)(n) * sizeof(*(src)))

261: /*MC
262:   PetscDeviceArrayZero - Zero memory in a device-aware manner

264:   Synopsis:
265: #include <petscdevice.h>
266:   PetscErrorCode PetscDeviceArrayZero(PetscDeviceContext dctx, void *ptr, size_t n)

268:   Not Collective, Asynchronous, Auto-dependency aware

270:   Input Parameters:
271: + dctx  - The `PetscDeviceContext` used to zero the memory
272: . ptr   - The pointer to the memory
273: - n     - The amount (in elements) to zero

275:   Level: beginner

277:   Notes:
278:   `ptr` must have been allocated using `PetscDeviceMalloc()` or `PetscDeviceCalloc()`.

280:   This uses the `sizeof()` of the memory type requested to determine the total memory to be
281:   zeroed, therefore you should not multiply the number of elements by the `sizeof()` the type\:

283: .vb
284:   PetscInt *ptr;

286:   // correct
287:   PetscDeviceArrayZero(dctx,ptr,n);

289:   // incorrect
290:   PetscDeviceArrayZero(dctx,ptr,n*sizeof(*ptr));
291: .ve

293:   See `PetscDeviceMemset()` for further discussion.

295: .N ASYNC_API

297: .seealso: `PetscDeviceMalloc()`, `PetscDeviceCalloc()`, `PetscDeviceFree()`,
298: `PetscDeviceArrayCopy()`, `PetscDeviceMemset()`
299: M*/
300: #define PetscDeviceArrayZero(dctx, ptr, n) PetscDeviceMemset((dctx), (ptr), 0, (size_t)(n) * sizeof(*(ptr)))