17d0a6c19SBarry Smith 2e5c89e4eSSatish Balay /* 3e5c89e4eSSatish Balay Code that allows a user to dictate what malloc() PETSc uses. 4e5c89e4eSSatish Balay */ 5c6db04a5SJed Brown #include <petscsys.h> /*I "petscsys.h" I*/ 6e5c89e4eSSatish Balay #if defined(PETSC_HAVE_MALLOC_H) 7e5c89e4eSSatish Balay #include <malloc.h> 8e5c89e4eSSatish Balay #endif 9e5c89e4eSSatish Balay 10e5c89e4eSSatish Balay /* 11e5c89e4eSSatish Balay We want to make sure that all mallocs of double or complex numbers are complex aligned. 12e5c89e4eSSatish Balay 1) on systems with memalign() we call that routine to get an aligned memory location 13e5c89e4eSSatish Balay 2) on systems without memalign() we 14e5c89e4eSSatish Balay - allocate one sizeof(PetscScalar) extra space 15e5c89e4eSSatish Balay - we shift the pointer up slightly if needed to get PetscScalar aligned 160700a824SBarry Smith - if shifted we store at ptr[-1] the amount of shift (plus a classid) 17e5c89e4eSSatish Balay */ 180700a824SBarry Smith #define SHIFT_CLASSID 456123 19e5c89e4eSSatish Balay 20e5c89e4eSSatish Balay #undef __FUNCT__ 21e5c89e4eSSatish Balay #define __FUNCT__ "PetscMallocAlign" 22efca3c55SSatish Balay PetscErrorCode PetscMallocAlign(size_t mem,int line,const char func[],const char file[],void **result) 23e5c89e4eSSatish Balay { 24f0ba7cfcSLisandro Dalcin if (!mem) { *result = NULL; return 0; } 25e5c89e4eSSatish Balay #if defined(PETSC_HAVE_DOUBLE_ALIGN_MALLOC) && (PETSC_MEMALIGN == 8) 26e5c89e4eSSatish Balay *result = malloc(mem); 27e5c89e4eSSatish Balay #elif defined(PETSC_HAVE_MEMALIGN) 28e5c89e4eSSatish Balay *result = memalign(PETSC_MEMALIGN,mem); 29e5c89e4eSSatish Balay #else 30e5c89e4eSSatish Balay { 31e5c89e4eSSatish Balay /* 32e5c89e4eSSatish Balay malloc space for two extra chunks and shift ptr 1 + enough to get it PetscScalar aligned 33e5c89e4eSSatish Balay */ 34f0ba7cfcSLisandro Dalcin int *ptr = (int*)malloc(mem + 2*PETSC_MEMALIGN); 35e5c89e4eSSatish Balay if (ptr) { 36f0ba7cfcSLisandro Dalcin int shift = (int)(((PETSC_UINTPTR_T) ptr) % PETSC_MEMALIGN); 37e5c89e4eSSatish Balay shift = (2*PETSC_MEMALIGN - shift)/sizeof(int); 380700a824SBarry Smith ptr[shift-1] = shift + SHIFT_CLASSID; 39e5c89e4eSSatish Balay ptr += shift; 40e5c89e4eSSatish Balay *result = (void*)ptr; 41e5c89e4eSSatish Balay } else { 42f0ba7cfcSLisandro Dalcin *result = NULL; 43e5c89e4eSSatish Balay } 44e5c89e4eSSatish Balay } 45e5c89e4eSSatish Balay #endif 46f0ba7cfcSLisandro Dalcin if (!*result) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_MEM,PETSC_ERROR_INITIAL,"Memory requested %.0f",(PetscLogDouble)mem); 47e5c89e4eSSatish Balay return 0; 48e5c89e4eSSatish Balay } 49e5c89e4eSSatish Balay 50e5c89e4eSSatish Balay #undef __FUNCT__ 51e5c89e4eSSatish Balay #define __FUNCT__ "PetscFreeAlign" 52efca3c55SSatish Balay PetscErrorCode PetscFreeAlign(void *ptr,int line,const char func[],const char file[]) 53e5c89e4eSSatish Balay { 54f0ba7cfcSLisandro Dalcin if (!ptr) return 0; 55e5c89e4eSSatish Balay #if (!(defined(PETSC_HAVE_DOUBLE_ALIGN_MALLOC) && (PETSC_MEMALIGN == 8)) && !defined(PETSC_HAVE_MEMALIGN)) 56f0ba7cfcSLisandro Dalcin { 57e5c89e4eSSatish Balay /* 58e5c89e4eSSatish Balay Previous int tells us how many ints the pointer has been shifted from 59e5c89e4eSSatish Balay the original address provided by the system malloc(). 60e5c89e4eSSatish Balay */ 61f0ba7cfcSLisandro Dalcin int shift = *(((int*)ptr)-1) - SHIFT_CLASSID; 62efca3c55SSatish Balay if (shift > PETSC_MEMALIGN-1) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"Likely memory corruption in heap"); 63efca3c55SSatish Balay if (shift < 0) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"Likely memory corruption in heap"); 64e5c89e4eSSatish Balay ptr = (void*)(((int*)ptr) - shift); 65e5c89e4eSSatish Balay } 66f0ba7cfcSLisandro Dalcin #endif 67e5c89e4eSSatish Balay 68e5c89e4eSSatish Balay #if defined(PETSC_HAVE_FREE_RETURN_INT) 69e5c89e4eSSatish Balay int err = free(ptr); 70efca3c55SSatish Balay if (err) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"System free returned error %d\n",err); 71e5c89e4eSSatish Balay #else 72e5c89e4eSSatish Balay free(ptr); 73e5c89e4eSSatish Balay #endif 74e5c89e4eSSatish Balay return 0; 75e5c89e4eSSatish Balay } 76e5c89e4eSSatish Balay 773221ece2SMatthew G. Knepley #undef __FUNCT__ 783221ece2SMatthew G. Knepley #define __FUNCT__ "PetscReallocAlign" 793221ece2SMatthew G. Knepley PetscErrorCode PetscReallocAlign(size_t mem, int line, const char func[], const char file[], void **result) 803221ece2SMatthew G. Knepley { 81*c22f1541SToby Isaac PetscErrorCode ierr; 82*c22f1541SToby Isaac 83*c22f1541SToby Isaac if (!mem) { 84*c22f1541SToby Isaac ierr = PetscFreeAlign(*result, line, func, file); 85*c22f1541SToby Isaac if (ierr) return ierr; 86*c22f1541SToby Isaac *result = NULL; 87*c22f1541SToby Isaac return 0; 88*c22f1541SToby Isaac } 893221ece2SMatthew G. Knepley #if (!(defined(PETSC_HAVE_DOUBLE_ALIGN_MALLOC) && (PETSC_MEMALIGN == 8)) && !defined(PETSC_HAVE_MEMALIGN)) 903221ece2SMatthew G. Knepley { 913221ece2SMatthew G. Knepley /* 923221ece2SMatthew G. Knepley Previous int tells us how many ints the pointer has been shifted from 933221ece2SMatthew G. Knepley the original address provided by the system malloc(). 943221ece2SMatthew G. Knepley */ 953221ece2SMatthew G. Knepley int shift = *(((int*)*result)-1) - SHIFT_CLASSID; 963221ece2SMatthew G. Knepley if (shift > PETSC_MEMALIGN-1) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"Likely memory corruption in heap"); 973221ece2SMatthew G. Knepley if (shift < 0) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"Likely memory corruption in heap"); 983221ece2SMatthew G. Knepley *result = (void*)(((int*)*result) - shift); 993221ece2SMatthew G. Knepley } 1003221ece2SMatthew G. Knepley #endif 1013221ece2SMatthew G. Knepley 102*c22f1541SToby Isaac #if (defined(PETSC_HAVE_DOUBLE_ALIGN_MALLOC) && (PETSC_MEMALIGN == 8)) || defined(PETSC_HAVE_MEMALIGN) 1033221ece2SMatthew G. Knepley *result = realloc(*result, mem); 1043221ece2SMatthew G. Knepley #else 1053221ece2SMatthew G. Knepley { 1063221ece2SMatthew G. Knepley /* 1073221ece2SMatthew G. Knepley malloc space for two extra chunks and shift ptr 1 + enough to get it PetscScalar aligned 1083221ece2SMatthew G. Knepley */ 1093221ece2SMatthew G. Knepley int *ptr = (int *) realloc(*result, mem + 2*PETSC_MEMALIGN); 1103221ece2SMatthew G. Knepley if (ptr) { 1113221ece2SMatthew G. Knepley int shift = (int)(((PETSC_UINTPTR_T) ptr) % PETSC_MEMALIGN); 1123221ece2SMatthew G. Knepley shift = (2*PETSC_MEMALIGN - shift)/sizeof(int); 1133221ece2SMatthew G. Knepley ptr[shift-1] = shift + SHIFT_CLASSID; 1143221ece2SMatthew G. Knepley ptr += shift; 1153221ece2SMatthew G. Knepley *result = (void*)ptr; 1163221ece2SMatthew G. Knepley } else { 1173221ece2SMatthew G. Knepley *result = NULL; 1183221ece2SMatthew G. Knepley } 1193221ece2SMatthew G. Knepley } 1203221ece2SMatthew G. Knepley #endif 1213221ece2SMatthew G. Knepley if (!*result) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_MEM,PETSC_ERROR_INITIAL,"Memory requested %.0f",(PetscLogDouble)mem); 122*c22f1541SToby Isaac #if defined(PETSC_HAVE_MEMALIGN) 123*c22f1541SToby Isaac /* There are no standard guarantees that realloc() maintains the alignment of memalign(), so I think we have to 124*c22f1541SToby Isaac * realloc and, if the alignment is wrong, malloc/copy/free. */ 125*c22f1541SToby Isaac if (((size_t) (*result)) % PETSC_MEMALIGN) { 126*c22f1541SToby Isaac void *newResult; 127*c22f1541SToby Isaac 128*c22f1541SToby Isaac newResult = memalign(PETSC_MEMALIGN,mem); 129*c22f1541SToby Isaac if (!newResult) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_MEM,PETSC_ERROR_INITIAL,"Memory requested %.0f",(PetscLogDouble)mem); 130*c22f1541SToby Isaac ierr = PetscMemcpy(newResult,*result,mem); 131*c22f1541SToby Isaac if (ierr) return ierr; 132*c22f1541SToby Isaac #if defined(PETSC_HAVE_FREE_RETURN_INT) 133*c22f1541SToby Isaac { 134*c22f1541SToby Isaac int err = free(*result); 135*c22f1541SToby Isaac if (err) return PetscError(PETSC_COMM_SELF,line,func,file,PETSC_ERR_PLIB,PETSC_ERROR_INITIAL,"System free returned error %d\n",err); 136*c22f1541SToby Isaac } 137*c22f1541SToby Isaac #else 138*c22f1541SToby Isaac free(*result); 139*c22f1541SToby Isaac #endif 140*c22f1541SToby Isaac *result = newResult; 141*c22f1541SToby Isaac } 142*c22f1541SToby Isaac #endif 1433221ece2SMatthew G. Knepley return 0; 1443221ece2SMatthew G. Knepley } 1453221ece2SMatthew G. Knepley 146efca3c55SSatish Balay PetscErrorCode (*PetscTrMalloc)(size_t,int,const char[],const char[],void**) = PetscMallocAlign; 147efca3c55SSatish Balay PetscErrorCode (*PetscTrFree)(void*,int,const char[],const char[]) = PetscFreeAlign; 1483221ece2SMatthew G. Knepley PetscErrorCode (*PetscTrRealloc)(size_t,int,const char[],const char[],void**) = PetscReallocAlign; 149e5c89e4eSSatish Balay 150ace3abfcSBarry Smith PetscBool petscsetmallocvisited = PETSC_FALSE; 151e5c89e4eSSatish Balay 152e5c89e4eSSatish Balay #undef __FUNCT__ 1531d1a0024SBarry Smith #define __FUNCT__ "PetscMallocSet" 154e5c89e4eSSatish Balay /*@C 1551d1a0024SBarry Smith PetscMallocSet - Sets the routines used to do mallocs and frees. 156e5c89e4eSSatish Balay This routine MUST be called before PetscInitialize() and may be 157e5c89e4eSSatish Balay called only once. 158e5c89e4eSSatish Balay 159e5c89e4eSSatish Balay Not Collective 160e5c89e4eSSatish Balay 161e5c89e4eSSatish Balay Input Parameters: 162e5c89e4eSSatish Balay + malloc - the malloc routine 163e5c89e4eSSatish Balay - free - the free routine 164e5c89e4eSSatish Balay 165e5c89e4eSSatish Balay Level: developer 166e5c89e4eSSatish Balay 167e5c89e4eSSatish Balay Concepts: malloc 168e5c89e4eSSatish Balay Concepts: memory^allocation 169e5c89e4eSSatish Balay 170e5c89e4eSSatish Balay @*/ 171efca3c55SSatish Balay PetscErrorCode PetscMallocSet(PetscErrorCode (*imalloc)(size_t,int,const char[],const char[],void**), 172efca3c55SSatish Balay PetscErrorCode (*ifree)(void*,int,const char[],const char[])) 173e5c89e4eSSatish Balay { 174e5c89e4eSSatish Balay PetscFunctionBegin; 175e32f2f54SBarry Smith if (petscsetmallocvisited && (imalloc != PetscTrMalloc || ifree != PetscTrFree)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"cannot call multiple times"); 176e5c89e4eSSatish Balay PetscTrMalloc = imalloc; 177e5c89e4eSSatish Balay PetscTrFree = ifree; 178e5c89e4eSSatish Balay petscsetmallocvisited = PETSC_TRUE; 179e5c89e4eSSatish Balay PetscFunctionReturn(0); 180e5c89e4eSSatish Balay } 181e5c89e4eSSatish Balay 182e5c89e4eSSatish Balay #undef __FUNCT__ 1831d1a0024SBarry Smith #define __FUNCT__ "PetscMallocClear" 184e5c89e4eSSatish Balay /*@C 1851d1a0024SBarry Smith PetscMallocClear - Resets the routines used to do mallocs and frees to the 186e5c89e4eSSatish Balay defaults. 187e5c89e4eSSatish Balay 188e5c89e4eSSatish Balay Not Collective 189e5c89e4eSSatish Balay 190e5c89e4eSSatish Balay Level: developer 191e5c89e4eSSatish Balay 192e5c89e4eSSatish Balay Notes: 193e5c89e4eSSatish Balay In general one should never run a PETSc program with different malloc() and 194e5c89e4eSSatish Balay free() settings for different parts; this is because one NEVER wants to 195e5c89e4eSSatish Balay free() an address that was malloced by a different memory management system 196e5c89e4eSSatish Balay 197e5c89e4eSSatish Balay @*/ 1987087cfbeSBarry Smith PetscErrorCode PetscMallocClear(void) 199e5c89e4eSSatish Balay { 200e5c89e4eSSatish Balay PetscFunctionBegin; 201e5c89e4eSSatish Balay PetscTrMalloc = PetscMallocAlign; 202e5c89e4eSSatish Balay PetscTrFree = PetscFreeAlign; 203e5c89e4eSSatish Balay petscsetmallocvisited = PETSC_FALSE; 204e5c89e4eSSatish Balay PetscFunctionReturn(0); 205e5c89e4eSSatish Balay } 206b44d5720SBarry Smith 207b44d5720SBarry Smith #undef __FUNCT__ 208b44d5720SBarry Smith #define __FUNCT__ "PetscMemoryTrace" 209b44d5720SBarry Smith PetscErrorCode PetscMemoryTrace(const char label[]) 210b44d5720SBarry Smith { 211b44d5720SBarry Smith PetscErrorCode ierr; 212b44d5720SBarry Smith PetscLogDouble mem,mal; 213b44d5720SBarry Smith static PetscLogDouble oldmem = 0,oldmal = 0; 214b44d5720SBarry Smith 215b44d5720SBarry Smith PetscFunctionBegin; 216b44d5720SBarry Smith ierr = PetscMemoryGetCurrentUsage(&mem);CHKERRQ(ierr); 217b44d5720SBarry Smith ierr = PetscMallocGetCurrentUsage(&mal);CHKERRQ(ierr); 218b44d5720SBarry Smith 219b44d5720SBarry Smith ierr = PetscPrintf(PETSC_COMM_WORLD,"%s High water %8.3f MB increase %8.3f MB Current %8.3f MB increase %8.3f MB\n",label,mem*1e-6,(mem - oldmem)*1e-6,mal*1e-6,(mal - oldmal)*1e-6);CHKERRQ(ierr); 220b44d5720SBarry Smith oldmem = mem; 221b44d5720SBarry Smith oldmal = mal; 222b44d5720SBarry Smith PetscFunctionReturn(0); 223b44d5720SBarry Smith } 224