static int DLOOP_Dataloop_create_flattened_struct(DLOOP_Count count,
const int *blklens,
const MPI_Aint *disps,
const DLOOP_Type *oldtypes,
DLOOP_Dataloop **dlp_p,
MPI_Aint *dlsz_p,
int *dldepth_p,
int flag)
{
/* arbitrary types, convert to bytes and use indexed */
int i, err, nr_blks = 0;
DLOOP_Size *tmp_blklens;
MPI_Aint *tmp_disps; /* since we're calling another fn that takes
this type as an input parameter */
DLOOP_Offset bytes;
DLOOP_Segment *segp;
int first_ind;
DLOOP_Size last_ind;
segp = MPIR_Segment_alloc();
/* --BEGIN ERROR HANDLING-- */
if (!segp) {
return DLOOP_Dataloop_create_struct_memory_error();
}
/* --END ERROR HANDLING-- */
/* use segment code once to count contiguous regions */
for (i=0; i < count; i++)
{
int is_basic;
/* ignore type elements with a zero blklen */
if (blklens[i] == 0) continue;
is_basic = (DLOOP_Handle_hasloop_macro(oldtypes[i])) ? 0 : 1;
if (is_basic && (oldtypes[i] != MPI_LB &&
oldtypes[i] != MPI_UB))
{
nr_blks++;
}
else /* derived type; get a count of contig blocks */
{
DLOOP_Count tmp_nr_blks, sz;
DLOOP_Handle_get_size_macro(oldtypes[i], sz);
/* if the derived type has some data to contribute,
* add to flattened representation */
if (sz > 0) {
err = MPIR_Segment_init(NULL,
(DLOOP_Count) blklens[i],
oldtypes[i],
segp,
flag);
if (err) return err;
bytes = SEGMENT_IGNORE_LAST;
MPIR_Segment_count_contig_blocks(segp,
0,
&bytes,
&tmp_nr_blks);
nr_blks += tmp_nr_blks;
}
}
}
/* it's possible for us to get to this point only to realize that
* there isn't any data in this type. in that case do what we always
* do: store a simple contig of zero ints and call it done.
*/
if (nr_blks == 0) {
MPIR_Segment_free(segp);
err = MPIR_Dataloop_create_contiguous(0,
MPI_INT,
dlp_p,
dlsz_p,
dldepth_p,
flag);
return err;
}
nr_blks += 2; /* safety measure */
tmp_blklens = (DLOOP_Size *) DLOOP_Malloc(nr_blks * sizeof(DLOOP_Size), MPL_MEM_DATATYPE);
/* --BEGIN ERROR HANDLING-- */
if (!tmp_blklens) {
MPIR_Segment_free(segp);
return DLOOP_Dataloop_create_struct_memory_error();
}
/* --END ERROR HANDLING-- */
tmp_disps = (MPI_Aint *) DLOOP_Malloc(nr_blks * sizeof(MPI_Aint), MPL_MEM_DATATYPE);
/* --BEGIN ERROR HANDLING-- */
if (!tmp_disps) {
DLOOP_Free(tmp_blklens);
MPIR_Segment_free(segp);
return DLOOP_Dataloop_create_struct_memory_error();
}
/* --END ERROR HANDLING-- */
/* use segment code again to flatten the type */
first_ind = 0;
for (i=0; i < count; i++)
{
int is_basic;
DLOOP_Count sz = -1;
is_basic = (DLOOP_Handle_hasloop_macro(oldtypes[i])) ? 0 : 1;
if (!is_basic) DLOOP_Handle_get_size_macro(oldtypes[i], sz);
/* we're going to use the segment code to flatten the type.
* we put in our displacement as the buffer location, and use
* the blocklength as the count value to get N contiguous copies
* of the type.
*
* Note that we're going to get back values in bytes, so that will
* be our new element type.
*/
if (oldtypes[i] != MPI_UB &&
oldtypes[i] != MPI_LB &&
blklens[i] != 0 &&
(is_basic || sz > 0))
{
err = MPIR_Segment_init((char *) DLOOP_OFFSET_CAST_TO_VOID_PTR disps[i],
(DLOOP_Count) blklens[i],
oldtypes[i],
segp,
0 /* homogeneous */);
if (err) return err;
last_ind = nr_blks - first_ind;
bytes = SEGMENT_IGNORE_LAST;
MPIR_Segment_mpi_flatten(segp,
0,
&bytes,
&tmp_blklens[first_ind],
&tmp_disps[first_ind],
&last_ind);
if (err) return err;
first_ind += last_ind;
}
}
nr_blks = first_ind;
#if 0
if (MPL_DBG_SELECTED(MPIR_DBG_DATATYPE,VERBOSE)) {
MPL_DBG_OUT(MPIR_DBG_DATATYPE,"--- start of flattened type ---");
for (i=0; i < nr_blks; i++) {
MPL_DBG_OUT_FMT(MPIR_DBG_DATATYPE,(MPL_DBG_FDEST,
"a[%d] = (%d, " DLOOP_OFFSET_FMT_DEC_SPEC ")", i,
tmp_blklens[i], tmp_disps[i]));
}
MPL_DBG_OUT(MPIR_DBG_DATATYPE,"--- end of flattened type ---");
}
#endif
MPIR_Segment_free(segp);
err = MPIR_Dataloop_create_indexed(nr_blks,
tmp_blklens,
tmp_disps,
1, /* disp in bytes */
MPI_BYTE,
dlp_p,
dlsz_p,
dldepth_p,
flag);
DLOOP_Free(tmp_blklens);
DLOOP_Free(tmp_disps);
return err;
}
int MPIDI_CH3U_Receive_data_found(MPIR_Request *rreq, void *buf, intptr_t *buflen, int *complete)
{
int dt_contig;
MPI_Aint dt_true_lb;
intptr_t userbuf_sz;
MPIR_Datatype * dt_ptr = NULL;
intptr_t data_sz;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3U_RECEIVE_DATA_FOUND);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3U_RECEIVE_DATA_FOUND);
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"posted request found");
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype,
dt_contig, userbuf_sz, dt_ptr, dt_true_lb);
if (rreq->dev.recv_data_sz <= userbuf_sz) {
data_sz = rreq->dev.recv_data_sz;
}
else {
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER,VERBOSE,(MPL_DBG_FDEST,
"receive buffer too small; message truncated, msg_sz=%" PRIdPTR ", userbuf_sz=%"
PRIdPTR,
rreq->dev.recv_data_sz, userbuf_sz));
rreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TRUNCATE,
"**truncate", "**truncate %d %d %d %d",
rreq->status.MPI_SOURCE, rreq->status.MPI_TAG,
rreq->dev.recv_data_sz, userbuf_sz );
MPIR_STATUS_SET_COUNT(rreq->status, userbuf_sz);
data_sz = userbuf_sz;
}
if (dt_contig && data_sz == rreq->dev.recv_data_sz)
{
/* user buffer is contiguous and large enough to store the
entire message. However, we haven't yet *read* the data
(this code describes how to read the data into the destination) */
/* if all of the data has already been received, unpack it
now, otherwise build an iov and let the channel unpack */
if (*buflen >= data_sz)
{
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"Copying contiguous data to user buffer");
/* copy data out of the receive buffer */
if (rreq->dev.drop_data == FALSE) {
MPIR_Memcpy((char*)(rreq->dev.user_buf) + dt_true_lb, buf, data_sz);
}
*buflen = data_sz;
*complete = TRUE;
}
else
{
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"IOV loaded for contiguous read");
rreq->dev.iov[0].MPL_IOV_BUF =
(MPL_IOV_BUF_CAST)((char*)(rreq->dev.user_buf) + dt_true_lb);
rreq->dev.iov[0].MPL_IOV_LEN = data_sz;
rreq->dev.iov_count = 1;
*buflen = 0;
*complete = FALSE;
}
/* Trigger OnFinal when receiving the last segment */
rreq->dev.OnDataAvail = rreq->dev.OnFinal;
}
else {
/* user buffer is not contiguous or is too small to hold
the entire message */
rreq->dev.segment_ptr = MPIR_Segment_alloc( );
MPIR_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPIR_Segment_alloc");
MPIR_Segment_init(rreq->dev.user_buf, rreq->dev.user_count,
rreq->dev.datatype, rreq->dev.segment_ptr);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
/* if all of the data has already been received, and the
message is not truncated, unpack it now, otherwise build an
iov and let the channel unpack */
if (data_sz == rreq->dev.recv_data_sz && *buflen >= data_sz)
{
intptr_t last;
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"Copying noncontiguous data to user buffer");
last = data_sz;
MPIR_Segment_unpack(rreq->dev.segment_ptr, rreq->dev.segment_first,
&last, buf);
/* --BEGIN ERROR HANDLING-- */
if (last != data_sz)
{
/* If the data can't be unpacked, the we have a
mismatch between the datatype and the amount of
data received. Throw away received data. */
MPIR_ERR_SET(rreq->status.MPI_ERROR, MPI_ERR_TYPE, "**dtypemismatch");
MPIR_STATUS_SET_COUNT(rreq->status, rreq->dev.segment_first);
*buflen = data_sz;
*complete = TRUE;
/* FIXME: Set OnDataAvail to 0? If not, why not? */
goto fn_exit;
}
/* --END ERROR HANDLING-- */
*buflen = data_sz;
/* Trigger OnFinal when receiving the last segment */
rreq->dev.OnDataAvail = rreq->dev.OnFinal;
*complete = TRUE;
}
else
{
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"IOV loaded for non-contiguous read");
mpi_errno = MPIDI_CH3U_Request_load_recv_iov(rreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_SETFATALANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|loadrecviov");
}
*buflen = 0;
*complete = FALSE;
}
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3U_RECEIVE_DATA_FOUND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3U_Post_data_receive_found(MPIR_Request * rreq)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPI_Aint dt_true_lb;
intptr_t userbuf_sz;
MPIR_Datatype * dt_ptr = NULL;
intptr_t data_sz;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_FOUND);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_FOUND);
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"posted request found");
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype,
dt_contig, userbuf_sz, dt_ptr, dt_true_lb);
if (rreq->dev.recv_data_sz <= userbuf_sz) {
data_sz = rreq->dev.recv_data_sz;
}
else {
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER,VERBOSE,(MPL_DBG_FDEST,
"receive buffer too small; message truncated, msg_sz=%" PRIdPTR ", userbuf_sz=%"
PRIdPTR,
rreq->dev.recv_data_sz, userbuf_sz));
rreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TRUNCATE,
"**truncate", "**truncate %d %d %d %d",
rreq->status.MPI_SOURCE, rreq->status.MPI_TAG,
rreq->dev.recv_data_sz, userbuf_sz );
MPIR_STATUS_SET_COUNT(rreq->status, userbuf_sz);
data_sz = userbuf_sz;
}
if (dt_contig && data_sz == rreq->dev.recv_data_sz)
{
/* user buffer is contiguous and large enough to store the
entire message. However, we haven't yet *read* the data
(this code describes how to read the data into the destination) */
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"IOV loaded for contiguous read");
rreq->dev.iov[0].MPL_IOV_BUF =
(MPL_IOV_BUF_CAST)((char*)(rreq->dev.user_buf) + dt_true_lb);
rreq->dev.iov[0].MPL_IOV_LEN = data_sz;
rreq->dev.iov_count = 1;
/* FIXME: We want to set the OnDataAvail to the appropriate
function, which depends on whether this is an RMA
request or a pt-to-pt request. */
rreq->dev.OnDataAvail = 0;
}
else {
/* user buffer is not contiguous or is too small to hold
the entire message */
MPL_DBG_MSG(MPIDI_CH3_DBG_OTHER,VERBOSE,"IOV loaded for non-contiguous read");
rreq->dev.segment_ptr = MPIR_Segment_alloc( );
MPIR_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPIR_Segment_alloc");
MPIR_Segment_init(rreq->dev.user_buf, rreq->dev.user_count,
rreq->dev.datatype, rreq->dev.segment_ptr);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
mpi_errno = MPIDI_CH3U_Request_load_recv_iov(rreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_SETFATALANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**ch3|loadrecviov");
}
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_FOUND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Unpack_impl(const void *inbuf, MPI_Aint insize, MPI_Aint * position,
void *outbuf, int outcount, MPI_Datatype datatype)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPIR_Segment *segp;
int contig;
MPI_Aint dt_true_lb;
MPI_Aint data_sz;
if (insize == 0)
goto fn_exit;
/* Handle contig case quickly */
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN) {
contig = TRUE;
dt_true_lb = 0;
data_sz = outcount * MPIR_Datatype_get_basic_size(datatype);
} else {
MPIR_Datatype *dt_ptr;
MPIR_Datatype_get_ptr(datatype, dt_ptr);
MPIR_Datatype_is_contig(datatype, &contig);
dt_true_lb = dt_ptr->true_lb;
data_sz = outcount * dt_ptr->size;
}
if (contig) {
MPIR_Memcpy((char *) outbuf + dt_true_lb, (char *) inbuf + *position, data_sz);
*position = (int) ((MPI_Aint) * position + data_sz);
goto fn_exit;
}
/* non-contig case */
segp = MPIR_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(segp == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s",
"MPIR_Segment_alloc");
mpi_errno = MPIR_Segment_init(outbuf, outcount, datatype, segp);
MPIR_Assert(mpi_errno == MPI_SUCCESS);
/* NOTE: the use of buffer values and positions in MPI_Unpack and in
* MPIR_Segment_unpack are quite different. See code or docs or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPIR_Ensure_Aint_fits_in_pointer((MPIR_VOID_PTR_CAST_TO_MPI_AINT inbuf) +
(MPI_Aint) * position);
MPIR_Segment_unpack(segp, first, &last, (void *) ((char *) inbuf + *position));
/* Ensure that calculation fits into an int datatype. */
MPIR_Ensure_Aint_fits_in_int((MPI_Aint) * position + last);
*position = (int) ((MPI_Aint) * position + last);
MPIR_Segment_free(segp);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
