xmp_coarray_fjrdma.c File Reference

#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>
#include <assert.h>
#include <string.h>
#include "mpi.h"
#include "mpi-ext.h"
#include "xmp_internal.h"

Include dependency graph for xmp_coarray_fjrdma.c:

Macros
#define	_XMP_FJRDMA_MAX_SIZE   16777212
#define	_XMP_FJRDMA_MAX_MEMID   511
#define	_XMP_FJRDMA_MAX_MPUT   1993
#define	_XMP_FJRDMA_MAX_MGET   100 /** This value is trial */
#define	_XMP_FJRDMA_MAX_PUT   60 /** This value is trial */
#define	_XMP_FJRDMA_MAX_GET   60 /** This value is trial */
#define	_XMP_COARRAY_SEND_NIC_TMP_0   FJMPI_RDMA_LOCAL_NIC0
#define	_XMP_COARRAY_SEND_NIC_TMP_1   FJMPI_RDMA_LOCAL_NIC1
#define	_XMP_COARRAY_SEND_NIC_TMP_2   FJMPI_RDMA_LOCAL_NIC2
#define	_XMP_COARRAY_SEND_NIC_TMP_3   FJMPI_RDMA_LOCAL_NIC3
#define	_XMP_COARRAY_FLAG_NIC_TMP_0   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC0)
#define	_XMP_COARRAY_FLAG_NIC_TMP_1   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC1)
#define	_XMP_COARRAY_FLAG_NIC_TMP_2   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC2)
#define	_XMP_COARRAY_FLAG_NIC_TMP_3   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC3)
#define	_XMP_COARRAY_FLAG_NIC_TMP_i0   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC0 | FJMPI_RDMA_IMMEDIATE_RETURN)
#define	_XMP_COARRAY_FLAG_NIC_TMP_i1   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC1 | FJMPI_RDMA_IMMEDIATE_RETURN)
#define	_XMP_COARRAY_FLAG_NIC_TMP_i2   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC2 | FJMPI_RDMA_IMMEDIATE_RETURN)
#define	_XMP_COARRAY_FLAG_NIC_TMP_i3   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC3 | FJMPI_RDMA_IMMEDIATE_RETURN)

Functions
void	_XMP_fjrdma_sync_memory_put ()
	End these variables are temporral. More...
void	_XMP_fjrdma_sync_memory_get ()
	Execute sync_memory for get operation. More...
void	_XMP_add_num_of_puts ()
	Add 1 to _num_of_puts. More...
void	_XMP_add_num_of_gets ()
	Add 1 to _num_of_gets. More...
void	_XMP_fjrdma_sync_memory ()
	Execute sync_memory. More...
void	_XMP_fjrdma_sync_all ()
	Execute sync_all. More...
void	_XMP_fjrdma_coarray_malloc (_XMP_coarray_t *coarray_desc, void **addr, const size_t coarray_size)
void	_XMP_fjrdma_regmem (_XMP_coarray_t *coarray_desc, void *addr, const size_t coarray_size)
void	_XMP_fjrdma_coarray_lastly_deallocate ()
	Deallocate memory region when calling _XMP_coarray_lastly_deallocate() More...
void	_XMP_fjrdma_contiguous_put (const int target_rank, const uint64_t dst_offset, const uint64_t src_offset, const _XMP_coarray_t *dst_desc, const _XMP_coarray_t *src_desc, const size_t dst_elmts, const size_t src_elmts, const size_t elmt_size)
void	_XMP_fjrdma_put (const int dst_contiguous, const int src_contiguous, const int target_rank, const int dst_dims, const int src_dims, const _XMP_array_section_t *dst_info, const _XMP_array_section_t *src_info, const _XMP_coarray_t *dst_desc, const _XMP_coarray_t *src_desc, void *src, const size_t dst_elmts, const size_t src_elmts)
void	_XMP_fjrdma_contiguous_get (const int target_rank, const _XMP_coarray_t *dst_desc, const _XMP_coarray_t *src_desc, const uint64_t dst_offset, const uint64_t src_offset, const size_t dst_elmts, const size_t src_elmts, const size_t elmt_size)
void	_XMP_fjrdma_get (const int src_contiguous, const int dst_contiguous, const int target_rank, const int src_dims, const int dst_dims, const _XMP_array_section_t *src_info, const _XMP_array_section_t *dst_info, const _XMP_coarray_t *src_desc, const _XMP_coarray_t *dst_desc, void *dst, const size_t src_elmts, const size_t dst_elmts)
void	_XMP_fjrdma_build_sync_images_table ()
	Build table and Initialize for sync images. More...
void	_XMP_fjrdma_sync_images (const int num, int *image_set, int *status)
	Execute sync images. More...

Variables
int	_XMP_flag_put_nb
	These variables are temporral. More...
int	_XMP_flag_put_nb_rr
int	_XMP_flag_put_nb_rr_i

Macro Definition Documentation

_XMP_COARRAY_FLAG_NIC_TMP_0

#define _XMP_COARRAY_FLAG_NIC_TMP_0   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC0)

_XMP_COARRAY_FLAG_NIC_TMP_1

#define _XMP_COARRAY_FLAG_NIC_TMP_1   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC1)

_XMP_COARRAY_FLAG_NIC_TMP_2

#define _XMP_COARRAY_FLAG_NIC_TMP_2   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC2)

_XMP_COARRAY_FLAG_NIC_TMP_3

#define _XMP_COARRAY_FLAG_NIC_TMP_3   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC3)

_XMP_COARRAY_FLAG_NIC_TMP_i0

#define _XMP_COARRAY_FLAG_NIC_TMP_i0   (FJMPI_RDMA_LOCAL_NIC0 | FJMPI_RDMA_REMOTE_NIC0 | FJMPI_RDMA_IMMEDIATE_RETURN)

_XMP_COARRAY_FLAG_NIC_TMP_i1

#define _XMP_COARRAY_FLAG_NIC_TMP_i1   (FJMPI_RDMA_LOCAL_NIC1 | FJMPI_RDMA_REMOTE_NIC1 | FJMPI_RDMA_IMMEDIATE_RETURN)

_XMP_COARRAY_FLAG_NIC_TMP_i2

#define _XMP_COARRAY_FLAG_NIC_TMP_i2   (FJMPI_RDMA_LOCAL_NIC2 | FJMPI_RDMA_REMOTE_NIC2 | FJMPI_RDMA_IMMEDIATE_RETURN)

_XMP_COARRAY_FLAG_NIC_TMP_i3

#define _XMP_COARRAY_FLAG_NIC_TMP_i3   (FJMPI_RDMA_LOCAL_NIC3 | FJMPI_RDMA_REMOTE_NIC3 | FJMPI_RDMA_IMMEDIATE_RETURN)

_XMP_COARRAY_SEND_NIC_TMP_0

#define _XMP_COARRAY_SEND_NIC_TMP_0   FJMPI_RDMA_LOCAL_NIC0

_XMP_COARRAY_SEND_NIC_TMP_1

#define _XMP_COARRAY_SEND_NIC_TMP_1   FJMPI_RDMA_LOCAL_NIC1

_XMP_COARRAY_SEND_NIC_TMP_2

#define _XMP_COARRAY_SEND_NIC_TMP_2   FJMPI_RDMA_LOCAL_NIC2

_XMP_COARRAY_SEND_NIC_TMP_3

#define _XMP_COARRAY_SEND_NIC_TMP_3   FJMPI_RDMA_LOCAL_NIC3

_XMP_FJRDMA_MAX_GET

#define _XMP_FJRDMA_MAX_GET   60 /** This value is trial */

_XMP_FJRDMA_MAX_MEMID

#define _XMP_FJRDMA_MAX_MEMID   511

_XMP_FJRDMA_MAX_MGET

#define _XMP_FJRDMA_MAX_MGET   100 /** This value is trial */

_XMP_FJRDMA_MAX_MPUT

#define _XMP_FJRDMA_MAX_MPUT   1993

_XMP_FJRDMA_MAX_PUT

#define _XMP_FJRDMA_MAX_PUT   60 /** This value is trial */

_XMP_FJRDMA_MAX_SIZE

#define _XMP_FJRDMA_MAX_SIZE   16777212

Function Documentation

_XMP_add_num_of_gets()

void _XMP_add_num_of_gets

(

)

Add 1 to _num_of_gets.

{
  _num_of_gets++;
  if(_num_of_gets > _XMP_FJRDMA_MAX_GET)
    _XMP_fjrdma_sync_memory_get();
}

Here is the caller graph for this function:

_XMP_add_num_of_puts()

void _XMP_add_num_of_puts

(

)

Add 1 to _num_of_puts.

{
  _num_of_puts++;
  if(_num_of_puts > _XMP_FJRDMA_MAX_PUT)
    _XMP_fjrdma_sync_memory_put();
}

Here is the caller graph for this function:

_XMP_fjrdma_build_sync_images_table()

void _XMP_fjrdma_build_sync_images_table

(

)

Build table and Initialize for sync images.

{
  _sync_images_table = malloc(sizeof(unsigned int) * _XMP_world_size);
 
  for(int i=0;i<_XMP_world_size;i++)
    _sync_images_table[i] = 0;
 
  double *token     = _XMP_alloc(sizeof(double));
  _local_rdma_addr  = FJMPI_Rdma_reg_mem(_XMP_SYNC_IMAGES_ID, token, sizeof(double));
  _remote_rdma_addr = _XMP_alloc(sizeof(uint64_t) * _XMP_world_size);
 
  // Obtain remote RDMA addresses
  MPI_Barrier(MPI_COMM_WORLD);
  for(int ncount=0,i=1; i<_XMP_world_size+1; ncount++,i++){
    int partner_rank = (_XMP_world_rank + _XMP_world_size - i) % _XMP_world_size;
    _remote_rdma_addr[partner_rank] = FJMPI_Rdma_get_remote_addr(partner_rank, _XMP_SYNC_IMAGES_ID);
 
    if(ncount > _XMP_INIT_RDMA_INTERVAL){
      MPI_Barrier(MPI_COMM_WORLD);
      ncount = 0;
    }
  }
}

Here is the caller graph for this function:

_XMP_fjrdma_coarray_lastly_deallocate()

void _XMP_fjrdma_coarray_lastly_deallocate

(

)

Deallocate memory region when calling _XMP_coarray_lastly_deallocate()

{
  if(_memid == _XMP_FJRDMA_START_MEMID) return;
 
  _memid--;
  FJMPI_Rdma_dereg_mem(_memid);
}

Here is the caller graph for this function:

_XMP_fjrdma_coarray_malloc()

void _XMP_fjrdma_coarray_malloc	(	_XMP_coarray_t *	coarray_desc,
		void **	addr,
		const size_t	coarray_size
	)

{
  *addr = _XMP_alloc(coarray_size);
  _XMP_fjrdma_regmem(coarray_desc, *addr, coarray_size);
}

Here is the call graph for this function:

_XMP_fjrdma_contiguous_get()

void _XMP_fjrdma_contiguous_get	(	const int	target_rank,
		const _XMP_coarray_t *	dst_desc,
		const _XMP_coarray_t *	src_desc,
		const uint64_t	dst_offset,
		const uint64_t	src_offset,
		const size_t	dst_elmts,
		const size_t	src_elmts,
		const size_t	elmt_size
	)

{
  size_t transfer_size = dst_elmts * elmt_size;
  _check_transfer_size(transfer_size);
 
  uint64_t raddr = (uint64_t)src_desc->addr[target_rank] + src_offset;
  uint64_t laddr = dst_desc->laddr + dst_offset;
  
  if(dst_elmts == src_elmts){
    _XMP_FJMPI_Rdma_get(target_rank, raddr, laddr, transfer_size);
    _XMP_fjrdma_sync_memory_get();
  }
  else if(src_elmts == 1){
    _XMP_FJMPI_Rdma_get(target_rank, raddr, laddr, elmt_size);
    _XMP_fjrdma_sync_memory_get();
 
    char *dst = dst_desc->real_addr + dst_offset;
    for(size_t i=1;i<dst_elmts;i++)
      memcpy(dst+i*elmt_size, dst, elmt_size);
  }
  else{
    _XMP_fatal("Coarray Error ! transfer size is wrong.\n");
  }
}

_XMP_fjrdma_contiguous_put()

void _XMP_fjrdma_contiguous_put	(	const int	target_rank,
		const uint64_t	dst_offset,
		const uint64_t	src_offset,
		const _XMP_coarray_t *	dst_desc,
		const _XMP_coarray_t *	src_desc,
		const size_t	dst_elmts,
		const size_t	src_elmts,
		const size_t	elmt_size
	)

{
  size_t transfer_size = dst_elmts * elmt_size;
  _check_transfer_size(transfer_size);
 
  uint64_t raddr = (uint64_t)dst_desc->addr[target_rank] + dst_offset;
  uint64_t laddr = src_desc->laddr + src_offset;
 
  if(dst_elmts == src_elmts){
    _XMP_FJMPI_Rdma_put(target_rank, raddr, laddr, transfer_size);
  }
  else if(src_elmts == 1){
    uint64_t raddrs[dst_elmts], laddrs[dst_elmts];
    size_t lengths[dst_elmts];
    for(size_t i=0;i<dst_elmts;i++) raddrs[i]  = raddr + i * elmt_size;
    for(size_t i=0;i<dst_elmts;i++) laddrs[i]  = laddr;
    for(size_t i=0;i<dst_elmts;i++) lengths[i] = elmt_size;
    _fjrdma_scalar_mput_do(target_rank, raddrs, laddrs, lengths, dst_elmts, elmt_size);
  }
  else{
    _XMP_fatal("Coarray Error ! transfer size is wrong.\n");
  }
 
  if(_XMP_flag_put_nb == false)
    _XMP_fjrdma_sync_memory_put();
}

_XMP_fjrdma_get()

void _XMP_fjrdma_get	(	const int	src_contiguous,
		const int	dst_contiguous,
		const int	target_rank,
		const int	src_dims,
		const int	dst_dims,
		const _XMP_array_section_t *	src_info,
		const _XMP_array_section_t *	dst_info,
		const _XMP_coarray_t *	src_desc,
		const _XMP_coarray_t *	dst_desc,
		void *	dst,
		const size_t	src_elmts,
		const size_t	dst_elmts
	)

{
  uint64_t dst_offset = (uint64_t)_XMP_get_offset(dst_info, dst_dims);
  uint64_t src_offset = (uint64_t)_XMP_get_offset(src_info, src_dims);
  size_t transfer_size = src_desc->elmt_size * src_elmts;
 
  _check_transfer_size(transfer_size);
 
  if(src_elmts == dst_elmts){
    if(dst_contiguous == _XMP_N_INT_TRUE && src_contiguous == _XMP_N_INT_TRUE){
      _fjrdma_contiguous_get(target_rank, dst_offset, src_offset, dst, dst_desc, src_desc, transfer_size);
    }
    else{
      _fjrdma_NON_contiguous_get(target_rank, dst_offset, src_offset, dst_info, src_info,
                                 dst, dst_desc, src_desc, dst_dims, src_dims, src_elmts);
    }
  }
  else{
    if(src_elmts == 1){
      _fjrdma_scalar_mget(target_rank, dst_offset, src_offset, dst_info, dst_dims, dst_desc, src_desc, (char *)dst, dst_elmts);
    }
    else{
      _XMP_fatal("Number of elements is invalid");
    }
  }
}

Here is the call graph for this function:

_XMP_fjrdma_put()

void _XMP_fjrdma_put	(	const int	dst_contiguous,
		const int	src_contiguous,
		const int	target_rank,
		const int	dst_dims,
		const int	src_dims,
		const _XMP_array_section_t *	dst_info,
		const _XMP_array_section_t *	src_info,
		const _XMP_coarray_t *	dst_desc,
		const _XMP_coarray_t *	src_desc,
		void *	src,
		const size_t	dst_elmts,
		const size_t	src_elmts
	)

{
  uint64_t dst_offset = (uint64_t)_XMP_get_offset(dst_info, dst_dims);
  uint64_t src_offset = (uint64_t)_XMP_get_offset(src_info, src_dims);
  size_t transfer_size = dst_desc->elmt_size * dst_elmts;
  _check_transfer_size(transfer_size);
 
  if(dst_elmts == src_elmts){
    if(dst_contiguous == _XMP_N_INT_TRUE && src_contiguous == _XMP_N_INT_TRUE){
      _fjrdma_contiguous_put(target_rank, dst_offset, src_offset, dst_desc, src_desc, src, transfer_size);
    }
    else{
      _fjrdma_NON_contiguous_put(target_rank, dst_offset, src_offset, dst_info, src_info, dst_dims, src_dims, 
                                 dst_desc, src_desc, src, dst_elmts);
    }
  }
  else{
    if(src_elmts == 1){
      _fjrdma_scalar_mput(target_rank, dst_offset, src_offset, dst_info, dst_dims, dst_desc, src_desc, 
                          src, dst_elmts);
    }
    else{
      _XMP_fatal("Number of elements is invalid");
    }
  }
}

Here is the call graph for this function:

_XMP_fjrdma_regmem()

void _XMP_fjrdma_regmem	(	_XMP_coarray_t *	coarray_desc,
		void *	addr,
		const size_t	coarray_size
	)

{
  uint64_t *each_addr = _XMP_alloc(sizeof(uint64_t) * _XMP_world_size);
  if(_memid == _XMP_FJRDMA_MAX_MEMID)
    _XMP_fatal("Too many coarrays. Number of coarrays is not more than 510.");
 
  coarray_desc->laddr = FJMPI_Rdma_reg_mem(_memid, addr, coarray_size);
 
  MPI_Barrier(MPI_COMM_WORLD);
  for(int ncount=0,i=1; i<_XMP_world_size+1; ncount++,i++){
    int partner_rank = (_XMP_world_rank + _XMP_world_size - i) % _XMP_world_size;
    each_addr[partner_rank] = FJMPI_Rdma_get_remote_addr(partner_rank, _memid);
    if(ncount > _XMP_INIT_RDMA_INTERVAL){
      MPI_Barrier(MPI_COMM_WORLD);
      ncount = 0;
    }
  }
 
  coarray_desc->real_addr = addr;
  coarray_desc->addr = (void *)each_addr;
  _memid++;
}

Here is the call graph for this function:

Here is the caller graph for this function:

_XMP_fjrdma_sync_all()

void _XMP_fjrdma_sync_all

(

)

Execute sync_all.

{
  if(_XMP_flag_put_nb)
    _XMP_fjrdma_sync_memory();
 
  //  _XMP_fjrdma_sync_memory();
  MPI_Barrier(MPI_COMM_WORLD);
}

Here is the call graph for this function:

Here is the caller graph for this function:

_XMP_fjrdma_sync_images()

void _XMP_fjrdma_sync_images	(	const int	num,
		int *	image_set,
		int *	status
	)

Execute sync images.

Parameters

[in]	num	number of nodes
[in]	*image_set	image set
[out]	status	status

{
  _XMP_fjrdma_sync_memory();
 
  if(num == 0){
    return;
  }
  else if(num < 0){
    fprintf(stderr, "Invalid value is used in xmp_sync_memory. The first argument is %d\n", num);
    _XMP_fatal_nomsg();
  }
 
  _notify_sync_images(num, image_set);
  _wait_sync_images(num, image_set);
 
  // Update table for post-processing
  for(int i=0;i<num;i++)
    _sync_images_table[image_set[i]]--;
}

Here is the call graph for this function:

Here is the caller graph for this function:

_XMP_fjrdma_sync_memory()

void _XMP_fjrdma_sync_memory

(

)

Execute sync_memory.

{
  if(_XMP_flag_put_nb)
    _XMP_fjrdma_sync_memory_put();
 
  //  _XMP_fjrdma_sync_memory_put();
  // _XMP_fjrdma_sync_memory_get don't need to be executed
}

Here is the call graph for this function:

Here is the caller graph for this function:

_XMP_fjrdma_sync_memory_get()

void _XMP_fjrdma_sync_memory_get

(

)

Execute sync_memory for get operation.

{
  while(_num_of_gets != 0)
    if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC, NULL) == FJMPI_RDMA_NOTICE)
      _num_of_gets--;
}

Here is the caller graph for this function:

_XMP_fjrdma_sync_memory_put()

void _XMP_fjrdma_sync_memory_put

(

)

End these variables are temporral.

Execute sync_memory for put operation

{
  if(_XMP_flag_put_nb_rr){
    while(1){
      if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC_TMP_0, NULL) == FJMPI_RDMA_NOTICE)
        _num_of_puts--;
      if(_num_of_puts == 0) break;
 
      if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC_TMP_1, NULL) == FJMPI_RDMA_NOTICE)
        _num_of_puts--;
      if(_num_of_puts == 0) break;
 
      if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC_TMP_2, NULL) == FJMPI_RDMA_NOTICE)
        _num_of_puts--;
      if(_num_of_puts == 0) break;
 
      if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC_TMP_3, NULL) == FJMPI_RDMA_NOTICE)
        _num_of_puts--;
      if(_num_of_puts == 0) break;
    }
  }
  else{
    while(_num_of_puts != 0)
      if(FJMPI_Rdma_poll_cq(_XMP_COARRAY_SEND_NIC, NULL) == FJMPI_RDMA_NOTICE)
        _num_of_puts--;
  }
}

Here is the caller graph for this function:

Variable Documentation

_XMP_flag_put_nb

int _XMP_flag_put_nb

These variables are temporral.

_XMP_flag_put_nb_rr

int _XMP_flag_put_nb_rr

_XMP_flag_put_nb_rr_i

int _XMP_flag_put_nb_rr_i