int OSPDI_GlobalBarrier()
{
int status = OSP_SUCCESS;
DCMF_Request_t request;
DCMF_Callback_t done_callback;
volatile int active;
OSPU_FUNC_ENTER();
active = 1;
done_callback.function = OSPDI_Generic_done;
done_callback.clientdata = (void *) &active;
status = DCMF_GlobalBarrier(&OSPD_GlobalBarrier_protocol,
&request,
done_callback);
OSPU_ERR_ABORT(status != DCMF_SUCCESS,
"DCMF_GlobalBarrier returned with an error");
OSPDI_Conditional_advance(active > 0);
fn_exit: OSPU_FUNC_EXIT();
return status;
fn_fail: goto fn_exit;
}
int OSPD_Get(int target, void* src, void* dst, int bytes)
{
int status = OSP_SUCCESS;
DCMF_Request_t request;
DCMF_Callback_t callback;
volatile int active;
unsigned src_disp, dst_disp;
OSPU_FUNC_ENTER();
OSPDI_CRITICAL_ENTER();
callback.function = OSPDI_Generic_done;
callback.clientdata = (void *) &active;
src_disp = (size_t) src - (size_t) OSPD_Membase_global[target];
dst_disp = (size_t) dst - (size_t) OSPD_Membase_global[OSPD_Process_info.my_rank];
active = 1;
status = DCMF_Get(&OSPD_Generic_get_protocol,
&request,
callback,
DCMF_RELAXED_CONSISTENCY,
target,
bytes,
&OSPD_Memregion_global[target],
&OSPD_Memregion_global[OSPD_Process_info.my_rank],
src_disp,
dst_disp);
OSPU_ERR_POP(status, "DCMF_Get returned with an error");
OSPDI_Conditional_advance(active > 0);
fn_exit:
OSPDI_CRITICAL_EXIT();
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSPD_PutAccV(int target,
OSP_iov_t *iov_ar,
int ar_len,
OSP_datatype_t osp_type,
void* scaling)
{
int status = OSP_SUCCESS;
OSPD_Handle_t *ospd_handle;
OSPU_FUNC_ENTER();
OSPDI_CRITICAL_ENTER();
ospd_handle = OSPDI_Get_handle();
OSPU_ERR_POP(status = (ospd_handle == NULL),
"OSPDI_Get_handle returned NULL in OSPD_PutAccS.\n");
status = OSPDI_Direct_putaccv(target,
iov_ar,
ar_len,
osp_type,
scaling,
ospd_handle);
OSPU_ERR_POP(status, "Direct putaccv function returned with an error \n");
OSPDI_Conditional_advance(ospd_handle->active > 0);
fn_exit:
OSPDI_Release_handle(ospd_handle);
OSPDI_CRITICAL_EXIT();
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSPDI_GlobalBcast(int root,
int count,
void *buffer)
{
int status = OSP_SUCCESS;
DCMF_Request_t request;
DCMF_Callback_t done_callback;
volatile unsigned gb_active = 0;
OSPU_FUNC_ENTER();
gb_active += 1;
done_callback.function = OSPDI_Generic_done;
done_callback.clientdata = (void *) &gb_active;
status = DCMF_GlobalBcast(&OSPD_GlobalBcast_protocol,
&request,
done_callback,
DCMF_SEQUENTIAL_CONSISTENCY,
root,
(char *) buffer,
count);
OSPU_ERR_POP(status != DCMF_SUCCESS,
"DCMF_GlobalBcast returned with error %d \n",
status);
OSPDI_Conditional_advance(gb_active > 0);
fn_exit:
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int OSPDI_GlobalAllreduce(int count,
OSP_reduce_op_t osp_op,
OSP_datatype_t osp_type,
void *in,
void *out)
{
int status = OSP_SUCCESS;
DCMF_CollectiveRequest_t ar_crequest;
DCMF_Callback_t done_callback;
DCMF_Op reduce_op;
DCMF_Dt datatype;
int bytes = 0;
void *in_abs = NULL;
volatile unsigned ga_active = 0;
OSPU_FUNC_ENTER();
switch (osp_op)
{
case OSP_SUM:
reduce_op = DCMF_SUM;
break;
case OSP_PROD:
reduce_op = DCMF_PROD;
break;
case OSP_MAX:
case OSP_MAXABS:
reduce_op = DCMF_MAX;
break;
case OSP_MIN:
case OSP_MINABS:
reduce_op = DCMF_MIN;
break;
case OSP_OR:
reduce_op = DCMF_LOR;
break;
default:
OSPU_ERR_POP(status != DCMF_SUCCESS, "Unsupported OSP_reduce_op \n");
break;
}
if (osp_op == OSP_MAXABS || osp_op == OSP_MINABS)
{
switch (osp_type)
{
case OSP_DOUBLE:
datatype = DCMF_DOUBLE;
bytes = count * sizeof(double);
status = OSPDI_Malloc(&in_abs, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS,
"OSPDI_Malloc returned error in OSPDI_GlobalAllreduce \n");
OSPDI_ABS(double, in, in_abs, count);
in = in_abs;
break;
case OSP_INT32:
datatype = DCMF_SIGNED_INT;
bytes = count * sizeof(int32_t);
status = OSPDI_Malloc(&in_abs, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS,
"OSPDI_Malloc returned error in OSPDI_GlobalAllreduce \n");
OSPDI_ABS(int32_t, in, in_abs, count);
in = in_abs;
break;
case OSP_INT64:
datatype = DCMF_SIGNED_LONG_LONG;
bytes = count * sizeof(int64_t);
status = OSPDI_Malloc(&in_abs, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS,
"OSPDI_Malloc returned error in OSPDI_GlobalAllreduce \n");
OSPDI_ABS(int64_t, in, in_abs, count);
in = in_abs;
break;
case OSP_UINT32:
datatype = DCMF_UNSIGNED_INT;
break;
case OSP_UINT64:
datatype = DCMF_UNSIGNED_LONG_LONG;
break;
case OSP_FLOAT:
datatype = DCMF_FLOAT;
bytes = count * sizeof(float);
status = OSPDI_Malloc(&in_abs, bytes);
OSPU_ERR_POP(status != OSP_SUCCESS,
"OSPDI_Malloc returned error in OSPDI_GlobalAllreduce \n");
OSPDI_ABS(float, in, in_abs, count);
in = in_abs;
break;
default:
OSPU_ERR_POP(status != DCMF_SUCCESS, "Unsupported OSP_datatype \n");
break;
}
}
else
{
switch (osp_type)
{
case OSP_DOUBLE:
datatype = DCMF_DOUBLE;
break;
case OSP_INT32:
datatype = DCMF_SIGNED_INT;
break;
case OSP_INT64:
datatype = DCMF_SIGNED_LONG_LONG;
break;
case OSP_UINT32:
datatype = DCMF_UNSIGNED_INT;
break;
case OSP_UINT64:
datatype = DCMF_UNSIGNED_LONG_LONG;
break;
case OSP_FLOAT:
datatype = DCMF_FLOAT;
break;
default:
OSPU_ERR_ABORT(status != DCMF_SUCCESS, "Unsupported OSP_datatype \n");
break;
}
}
ga_active += 1;
done_callback.function = OSPDI_Generic_done;
done_callback.clientdata = (void *) &ga_active;
status = DCMF_Allreduce(&OSPD_GlobalAllreduce_protocol,
&ar_crequest,
done_callback,
DCMF_SEQUENTIAL_CONSISTENCY,
&geometry,
(char *) in,
(char *) out,
count,
datatype,
reduce_op);
OSPDI_Conditional_advance(ga_active > 0);
fn_exit:
if (in_abs != NULL)
OSPDI_Free(in_abs);
OSPU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
