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 { 24*f0ba7cfcSLisandro 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 */ 34*f0ba7cfcSLisandro Dalcin int *ptr = (int*)malloc(mem + 2*PETSC_MEMALIGN); 35e5c89e4eSSatish Balay if (ptr) { 36*f0ba7cfcSLisandro 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 { 42*f0ba7cfcSLisandro Dalcin *result = NULL; 43e5c89e4eSSatish Balay } 44e5c89e4eSSatish Balay } 45e5c89e4eSSatish Balay #endif 46*f0ba7cfcSLisandro 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 { 54*f0ba7cfcSLisandro Dalcin if (!ptr) return 0; 55e5c89e4eSSatish Balay #if (!(defined(PETSC_HAVE_DOUBLE_ALIGN_MALLOC) && (PETSC_MEMALIGN == 8)) && !defined(PETSC_HAVE_MEMALIGN)) 56*f0ba7cfcSLisandro 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 */ 61*f0ba7cfcSLisandro 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 } 66*f0ba7cfcSLisandro 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 77efca3c55SSatish Balay PetscErrorCode (*PetscTrMalloc)(size_t,int,const char[],const char[],void**) = PetscMallocAlign; 78efca3c55SSatish Balay PetscErrorCode (*PetscTrFree)(void*,int,const char[],const char[]) = PetscFreeAlign; 79e5c89e4eSSatish Balay 80ace3abfcSBarry Smith PetscBool petscsetmallocvisited = PETSC_FALSE; 81e5c89e4eSSatish Balay 82e5c89e4eSSatish Balay #undef __FUNCT__ 831d1a0024SBarry Smith #define __FUNCT__ "PetscMallocSet" 84e5c89e4eSSatish Balay /*@C 851d1a0024SBarry Smith PetscMallocSet - Sets the routines used to do mallocs and frees. 86e5c89e4eSSatish Balay This routine MUST be called before PetscInitialize() and may be 87e5c89e4eSSatish Balay called only once. 88e5c89e4eSSatish Balay 89e5c89e4eSSatish Balay Not Collective 90e5c89e4eSSatish Balay 91e5c89e4eSSatish Balay Input Parameters: 92e5c89e4eSSatish Balay + malloc - the malloc routine 93e5c89e4eSSatish Balay - free - the free routine 94e5c89e4eSSatish Balay 95e5c89e4eSSatish Balay Level: developer 96e5c89e4eSSatish Balay 97e5c89e4eSSatish Balay Concepts: malloc 98e5c89e4eSSatish Balay Concepts: memory^allocation 99e5c89e4eSSatish Balay 100e5c89e4eSSatish Balay @*/ 101efca3c55SSatish Balay PetscErrorCode PetscMallocSet(PetscErrorCode (*imalloc)(size_t,int,const char[],const char[],void**), 102efca3c55SSatish Balay PetscErrorCode (*ifree)(void*,int,const char[],const char[])) 103e5c89e4eSSatish Balay { 104e5c89e4eSSatish Balay PetscFunctionBegin; 105e32f2f54SBarry Smith if (petscsetmallocvisited && (imalloc != PetscTrMalloc || ifree != PetscTrFree)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"cannot call multiple times"); 106e5c89e4eSSatish Balay PetscTrMalloc = imalloc; 107e5c89e4eSSatish Balay PetscTrFree = ifree; 108e5c89e4eSSatish Balay petscsetmallocvisited = PETSC_TRUE; 109e5c89e4eSSatish Balay PetscFunctionReturn(0); 110e5c89e4eSSatish Balay } 111e5c89e4eSSatish Balay 112e5c89e4eSSatish Balay #undef __FUNCT__ 1131d1a0024SBarry Smith #define __FUNCT__ "PetscMallocClear" 114e5c89e4eSSatish Balay /*@C 1151d1a0024SBarry Smith PetscMallocClear - Resets the routines used to do mallocs and frees to the 116e5c89e4eSSatish Balay defaults. 117e5c89e4eSSatish Balay 118e5c89e4eSSatish Balay Not Collective 119e5c89e4eSSatish Balay 120e5c89e4eSSatish Balay Level: developer 121e5c89e4eSSatish Balay 122e5c89e4eSSatish Balay Notes: 123e5c89e4eSSatish Balay In general one should never run a PETSc program with different malloc() and 124e5c89e4eSSatish Balay free() settings for different parts; this is because one NEVER wants to 125e5c89e4eSSatish Balay free() an address that was malloced by a different memory management system 126e5c89e4eSSatish Balay 127e5c89e4eSSatish Balay @*/ 1287087cfbeSBarry Smith PetscErrorCode PetscMallocClear(void) 129e5c89e4eSSatish Balay { 130e5c89e4eSSatish Balay PetscFunctionBegin; 131e5c89e4eSSatish Balay PetscTrMalloc = PetscMallocAlign; 132e5c89e4eSSatish Balay PetscTrFree = PetscFreeAlign; 133e5c89e4eSSatish Balay petscsetmallocvisited = PETSC_FALSE; 134e5c89e4eSSatish Balay PetscFunctionReturn(0); 135e5c89e4eSSatish Balay } 136b44d5720SBarry Smith 137b44d5720SBarry Smith #undef __FUNCT__ 138b44d5720SBarry Smith #define __FUNCT__ "PetscMemoryTrace" 139b44d5720SBarry Smith PetscErrorCode PetscMemoryTrace(const char label[]) 140b44d5720SBarry Smith { 141b44d5720SBarry Smith PetscErrorCode ierr; 142b44d5720SBarry Smith PetscLogDouble mem,mal; 143b44d5720SBarry Smith static PetscLogDouble oldmem = 0,oldmal = 0; 144b44d5720SBarry Smith 145b44d5720SBarry Smith PetscFunctionBegin; 146b44d5720SBarry Smith ierr = PetscMemoryGetCurrentUsage(&mem);CHKERRQ(ierr); 147b44d5720SBarry Smith ierr = PetscMallocGetCurrentUsage(&mal);CHKERRQ(ierr); 148b44d5720SBarry Smith 149b44d5720SBarry 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); 150b44d5720SBarry Smith oldmem = mem; 151b44d5720SBarry Smith oldmal = mal; 152b44d5720SBarry Smith PetscFunctionReturn(0); 153b44d5720SBarry Smith } 154