int A1_Flush(int proc)
{
int status = A1_SUCCESS;
A1U_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef A1_TAU_PROFILING
{
TAU_TRACE_SENDMSG (A1_TAU_TAG_FLUSH, proc, 8);
}
# endif
status = A1D_Flush(proc);
A1U_ERR_POP(status!=A1_SUCCESS, "A1D_Flush returned an error\n");
fn_exit:
A1U_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSP_NbGet(int target, void* src, void* dst, int bytes, OSP_handle_t osp_handle)
{
int status = OSP_SUCCESS;
int my_rank = OSPD_Process_id(OSP_GROUP_WORLD);
OSPU_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef OSP_TAU_PROFILING
{
TAU_TRACE_SENDMSG (OSP_TAU_TAG_NBGET, target, bytes);
}
# endif
if(target == my_rank && (bytes < ospu_settings.network_bypass_upper_limit_1d) )
{
status = OSPU_Get_memcpy(src, dst, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPU_Get_memcpy returned an error\n");
}
else
{
status = OSPD_NbGet(target, src, dst, bytes, osp_handle);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPD_NbGet returned an error\n");
}
fn_exit:
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int A1_NbPutAcc(int target,
void* source_ptr,
void* target_ptr,
int bytes,
A1_datatype_t a1_type,
void* scaling,
A1_handle_t a1_handle)
{
int status = A1_SUCCESS;
int my_rank = A1D_Process_id(A1_GROUP_WORLD);
A1U_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef A1_TAU_PROFILING
{
TAU_TRACE_SENDMSG (A1_TAU_TAG_NBPUTACC, target, bytes);
}
# endif
/* Bypass is ALWAYS better for accumulate; we do not test against threshold. */
if (target == my_rank && a1u_settings.network_bypass)
{
status = A1U_Acc_memcpy(source_ptr,
target_ptr,
bytes,
a1_type,
scaling);
A1U_ERR_POP(status != A1_SUCCESS, "A1U_Acc_memcpy returned an error\n");
}
else
{
status = A1D_NbPutAcc(target,
source_ptr,
target_ptr,
bytes,
a1_type,
scaling,
a1_handle);
A1U_ERR_POP((status!=A1_SUCCESS), "A1D_NbPutAcc returned error\n");
}
fn_exit: A1U_FUNC_EXIT();
return status;
fn_fail: goto fn_exit;
}
int OSP_NbPutAccV(int target,
OSP_iov_t *iov_ar,
int ar_len,
OSP_datatype_t osp_type,
void* scaling,
OSP_handle_t osp_handle)
{
int status = OSP_SUCCESS;
int my_rank = OSPD_Process_id(OSP_GROUP_WORLD);
OSPU_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef OSP_TAU_PROFILING
{
int i, total_bytes = 0;
for (i = 0; i < ar_len; i++)
total_bytes += iov_ar[i].ptr_array_len * iov_ar[i].bytes;
TAU_TRACE_SENDMSG (OSP_TAU_TAG_NBPUTACCV, target, total_bytes);
}
# endif
/* Bypass is ALWAYS better for accumulate; we do not test against threshold. */
if (target == my_rank && ospu_settings.network_bypass)
{
status = OSPU_AccV_memcpy(iov_ar, ar_len, osp_type, scaling);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPU_AccV_memcpy returned an error\n");
}
else
{
status = OSPD_NbPutAccV(target,
iov_ar,
ar_len,
osp_type,
scaling,
osp_handle);
OSPU_ERR_POP(status, "OSPD_NbPutAccV returned error\n");
}
fn_exit: OSPU_FUNC_EXIT();
return status;
fn_fail: goto fn_exit;
}
int OSP_NbPut(int target, void* src, void* dst, int bytes, OSP_handle_t osp_handle)
{
int status = OSP_SUCCESS;
int my_rank = OSPD_Process_id(OSP_GROUP_WORLD);
OSPU_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef OSP_TAU_PROFILING
{
TAU_TRACE_SENDMSG (OSP_TAU_TAG_NBPUT, target, bytes);
}
# endif
/* Not sure if what is the right strategy for bypass. OSPU_*_memcpy are blocking
* but the overhead of going into DCMF_Put is likely not worth the savings
* from said call being non-blocking. This is especially true under heavy load
* since we have determined that DMA vs. memcpy turns over when the NIC is getting
* hammered.
*/
if(target == my_rank && (bytes < ospu_settings.network_bypass_upper_limit_1d) )
{
status = OSPU_Put_memcpy(src, dst, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPU_Put_memcpy returned an error\n");
}
else
{
status = OSPD_NbPut(target, src, dst, bytes, osp_handle);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPD_NbPut returned an error\n");
}
fn_exit:
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int A1_PutModV(int target,
A1_iov_t *iov_ar,
int ar_len,
A1_reduce_op_t a1_op,
A1_datatype_t a1_type)
{
int status = A1_SUCCESS;
int my_rank = A1D_Process_id(A1_GROUP_WORLD);
A1U_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef A1_TAU_PROFILING
{
int i, total_bytes = 0;
for (i = 0; i < ar_len; i++)
total_bytes += iov_ar[i].ptr_array_len * iov_ar[i].bytes;
TAU_TRACE_SENDMSG (A1_TAU_TAG_PUTMODV, target, total_bytes);
}
# endif
/* Bypass is ALWAYS better for accumulate; we do not test against threshold. */
if (target == my_rank && a1u_settings.network_bypass)
{
status = A1U_ModV_memcpy(iov_ar, ar_len, a1_op, a1_type);
A1U_ERR_POP(status != A1_SUCCESS, "A1U_ModV_memcpy returned an error\n");
}
else
{
status = A1D_PutModV(target, iov_ar, ar_len, a1_op, a1_type);
A1U_ERR_POP(status, "A1D_PutModV returned error\n");
}
fn_exit: A1U_FUNC_EXIT();
return status;
fn_fail: goto fn_exit;
}
int A1_Rmw(int target,
void* source_ptr_in,
void* source_ptr_out,
void* target_ptr,
int bytes,
A1_atomic_op_t op,
A1_datatype_t a1_type)
{
int status = A1_SUCCESS;
A1U_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef A1_TAU_PROFILING
{
TAU_TRACE_SENDMSG (A1_TAU_TAG_RMW, target, bytes);
}
# endif
status = A1D_Rmw(target,
source_ptr_in,
source_ptr_out,
target_ptr,
bytes,
op,
a1_type);
A1U_ERR_POP(status!=A1_SUCCESS, "A1D_Rmw returned an error\n");
fn_exit:
A1U_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSP_Rmw(int target,
void* source_ptr_in,
void* source_ptr_out,
void* target_ptr,
int bytes,
OSP_atomic_op_t op,
OSP_datatype_t osp_type)
{
int status = OSP_SUCCESS;
OSPU_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef OSP_TAU_PROFILING
{
TAU_TRACE_SENDMSG (OSP_TAU_TAG_RMW, target, bytes);
}
# endif
status = OSPD_Rmw(target,
source_ptr_in,
source_ptr_out,
target_ptr,
bytes,
op,
osp_type);
OSPU_ERR_POP(status!=OSP_SUCCESS, "OSPD_Rmw returned an error\n");
fn_exit:
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSP_NbPutAccS(int target,
int stride_level,
int *block_sizes,
void* source_ptr,
int *src_stride_ar,
void* target_ptr,
int *trg_stride_ar,
OSP_datatype_t osp_type,
void* scaling,
OSP_handle_t osp_handle)
{
int status = OSP_SUCCESS;
int my_rank = OSPD_Process_id(OSP_GROUP_WORLD);
OSPU_FUNC_ENTER();
# ifdef HAVE_ERROR_CHECKING
# endif
# ifdef OSP_TAU_PROFILING
{
int i, bytes = 1;
for (i = 0; i <= stride_levels; i++) total_bytes *= count[i];
TAU_TRACE_SENDMSG (OSP_TAU_TAG_NBPUTACCS, target, total_bytes);
}
# endif
/* Bypass is ALWAYS better for accumulate; we do not test against threshold. */
if (target == my_rank && ospu_settings.network_bypass)
{
status = OSPU_AccS_local(stride_level,
block_sizes,
source_ptr,
src_stride_ar,
target_ptr,
trg_stride_ar,
osp_type,
scaling);
OSPU_ERR_POP(status != OSP_SUCCESS, "OSPU_AccS_local returned an error\n");
}
else
{
status = OSPI_Recursive_PutAcc(target,
stride_level,
block_sizes,
source_ptr,
src_stride_ar,
target_ptr,
trg_stride_ar,
osp_type,
scaling,
osp_handle);
OSPU_ERR_POP(status!=OSP_SUCCESS, "OSPI_Recursive_PutAcc returned error\n");
}
fn_exit:
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
