xmp_reduce_hybrid.c File Reference

#include <cuda_runtime.h>
#include <xmp_internal.h>
#include "tca-api.h"
#include <math.h>

Include dependency graph for xmp_reduce_hybrid.c:

Data Structures
struct	_XMP_tca_coll_info_type

Macros
#define	_XMP_TCA_SYNC_MARK_SIZE   sizeof(unsigned long)
#define	_XMP_TCA_CACHE_ALIGNED_STRIDE   64
#define	_XMP_TCA_PIO_SYNC_MARK   255
#define	_XMP_TCA_COLL_MAX   64
#define	_XMP_TCA_ALLREDUCE_TAG   0x100
#define	_XMP_TCA_DEVICE_TO_HOST_WAIT_SLOT   0
#define	_XMP_TCA_HOST_TO_DEVICE_WAIT_SLOT   1
#define	_XMP_TCA_ALLREDUCE_TCACOPY_LIMIT   8
#define	CUDA_CHECK(cuda_call)
#define	OP_FUNC_3OP(name, op, type_name, type)
#define	TCA_TYPE_FUNCTIONS(name, type)

Typedefs
typedef struct _XMP_tca_coll_info_type	_XMP_tca_coll_info_t
typedef void(*	tca_op_func_3op_handler_t) (void *, void *, void *, int)
typedef tca_op_func_3op_handler_t	tca_op_fn_3op_t

Enumerations
enum	{   TCA_OP_INT8, TCA_OP_UINT8, TCA_OP_INT16, TCA_OP_UINT16,   TCA_OP_INT32, TCA_OP_UINT32, TCA_OP_INT64, TCA_OP_UINT64,   TCA_OP_FLOAT, TCA_OP_DOUBLE, TCA_OP_LONG_DOUBLE, TCA_OP_TYPE_MAX }

Functions
void	_XMP_reduce_hybrid_NODES_ENTIRE (_XMP_nodes_t *nodes, void *addr, int count, int datatype, int op)
void	_XMP_reduce_hybrid_CLAUSE (void *data_addr, int count, int datatype, int op)

Variables
_XMP_tca_coll_info_t	coll_info
int	_XMP_tca_coll_info_flag

Macro Definition Documentation

_XMP_TCA_ALLREDUCE_TAG

#define _XMP_TCA_ALLREDUCE_TAG   0x100

_XMP_TCA_ALLREDUCE_TCACOPY_LIMIT

#define _XMP_TCA_ALLREDUCE_TCACOPY_LIMIT   8

_XMP_TCA_CACHE_ALIGNED_STRIDE

#define _XMP_TCA_CACHE_ALIGNED_STRIDE   64

_XMP_TCA_COLL_MAX

#define _XMP_TCA_COLL_MAX   64

_XMP_TCA_DEVICE_TO_HOST_WAIT_SLOT

#define _XMP_TCA_DEVICE_TO_HOST_WAIT_SLOT   0

_XMP_TCA_HOST_TO_DEVICE_WAIT_SLOT

#define _XMP_TCA_HOST_TO_DEVICE_WAIT_SLOT   1

_XMP_TCA_PIO_SYNC_MARK

#define _XMP_TCA_PIO_SYNC_MARK   255

_XMP_TCA_SYNC_MARK_SIZE

#define _XMP_TCA_SYNC_MARK_SIZE   sizeof(unsigned long)

CUDA_CHECK

#define CUDA_CHECK

(

cuda_call

)

Value:

    do {                                      \
    cudaError_t status = cuda_call;                                     \
    if(status != cudaSuccess) {                                         \
      fprintf(stderr, "CUDA error in file '%s' in line %i : %s.\n",     \
              __FILE__, __LINE__, cudaGetErrorString(status) );         \
      exit(EXIT_FAILURE);                                               \
    }                                                                   \
  } while (0)

OP_FUNC_3OP

#define OP_FUNC_3OP	(		name,
			op,
			type_name,
			type
	)

Value:

  static inline void tca_op_func_3op_##name##_##type_name(void *dst, void *src0, void *src1, int count) { \
    int i;                                                              \
    type *s0 = (type *)src0;                                            \
    type *s1 = (type *)src1;                                            \
    type *d = (type *)dst;                                              \
    for(i = 0; i < count; i++) {                                        \
      *(d++) = *(s0++) op *(s1++);                                      \
    }                                                                   \
  }

TCA_TYPE_FUNCTIONS

#define TCA_TYPE_FUNCTIONS	(		name,
			type
	)

Value:

  [TCA_OP_INT8]        = tca_op_func_##type##_##name##_int8,            \
    [TCA_OP_UINT8]       = tca_op_func_##type##_##name##_uint8,         \
    [TCA_OP_INT16]       = tca_op_func_##type##_##name##_int16,         \
    [TCA_OP_UINT16]      = tca_op_func_##type##_##name##_uint16,        \
    [TCA_OP_INT32]       = tca_op_func_##type##_##name##_int32,         \
    [TCA_OP_UINT32]      = tca_op_func_##type##_##name##_uint32,        \
    [TCA_OP_INT64]       = tca_op_func_##type##_##name##_int64,         \
    [TCA_OP_UINT64]      = tca_op_func_##type##_##name##_uint64,        \
    [TCA_OP_FLOAT]       = tca_op_func_##type##_##name##_float,         \
    [TCA_OP_DOUBLE]      = tca_op_func_##type##_##name##_double,        \
    [TCA_OP_LONG_DOUBLE] = tca_op_func_##type##_##name##_long_double    \

Typedef Documentation

_XMP_tca_coll_info_t

typedef struct _XMP_tca_coll_info_type _XMP_tca_coll_info_t

tca_op_fn_3op_t

typedef tca_op_func_3op_handler_t tca_op_fn_3op_t

tca_op_func_3op_handler_t

typedef void(* tca_op_func_3op_handler_t) (void *, void *, void *, int)

Enumeration Type Documentation

anonymous enum

anonymous enum

Enumerator
TCA_OP_INT8
TCA_OP_UINT8
TCA_OP_INT16
TCA_OP_UINT16
TCA_OP_INT32
TCA_OP_UINT32
TCA_OP_INT64
TCA_OP_UINT64
TCA_OP_FLOAT
TCA_OP_DOUBLE
TCA_OP_LONG_DOUBLE
TCA_OP_TYPE_MAX

     {
  TCA_OP_INT8,
  TCA_OP_UINT8,
  TCA_OP_INT16,
  TCA_OP_UINT16,
  TCA_OP_INT32,
  TCA_OP_UINT32,
  TCA_OP_INT64,
  TCA_OP_UINT64,
  TCA_OP_FLOAT,
  TCA_OP_DOUBLE,
  TCA_OP_LONG_DOUBLE,
  TCA_OP_TYPE_MAX,
};

Function Documentation

_XMP_reduce_hybrid_CLAUSE()

void _XMP_reduce_hybrid_CLAUSE	(	void *	data_addr,
		int	count,
		int	datatype,
		int	op
	)

{
  // Not implemented
  _XMP_fatal("_XMP_reduce_hybrid_CLAUSE is not implemented.");
}

Here is the call graph for this function:

_XMP_reduce_hybrid_NODES_ENTIRE()

void _XMP_reduce_hybrid_NODES_ENTIRE	(	_XMP_nodes_t *	nodes,
		void *	addr,
		int	count,
		int	datatype,
		int	op
	)

{
  if (count == 0) {
    return; // FIXME not good implementation
  }
  if (!nodes->is_member) {
    return;
  }
  if (_XMP_tca_coll_info_flag) {
    init_coll_info();
  }
 
  MPI_Comm mpi_comm = *((MPI_Comm *)nodes->comm);
 
  int id = get_coll_id(dev_addr, count, datatype, op, mpi_comm);
  if (!coll_info.flag[id]) {
    _XMP_reduce_init_hybrid(dev_addr, count, datatype, op, mpi_comm, id);
  }
 
  _XMP_reduce_do_hybrid(dev_addr, count, datatype, op, mpi_comm, id);
}

Variable Documentation

_XMP_tca_coll_info_flag

int _XMP_tca_coll_info_flag

coll_info

_XMP_tca_coll_info_t coll_info