/** Prepare a set of buffers for use with an accumulate operation. The
* returned set of buffers is guaranteed to be in private space and scaled.
* Copies will be made if needed, the result should be completed by finish.
*
* @param[in] orig_bufs Original set of buffers.
* @param[out] new_bufs Pointer to the set of private buffers.
* @param[in] count Number of entries in the buffer list.
* @param[in] size The size of the buffers (all are of the same size).
* @param[in] datatype The type of the buffer.
* @param[in] scale Scaling constant to apply to each buffer.
* @return Number of buffers that were moved.
*/
int ARMCII_Buf_prepare_acc_vec(void **orig_bufs, void ***new_bufs_ptr, int count, int size,
int datatype, void *scale) {
void **new_bufs;
int i, scaled, num_moved = 0;
new_bufs = malloc(count*sizeof(void*));
ARMCII_Assert(new_bufs != NULL);
scaled = ARMCII_Buf_acc_is_scaled(datatype, scale);
for (i = 0; i < count; i++) {
gmr_t *mreg = NULL;
// Check if the source buffer is within a shared region.
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
mreg = gmr_lookup(orig_bufs[i], ARMCI_GROUP_WORLD.rank);
if (scaled) {
MPI_Alloc_mem(size, MPI_INFO_NULL, &new_bufs[i]);
ARMCII_Assert(new_bufs[i] != NULL);
// Lock if needed so we can directly access the buffer
if (mreg != NULL)
gmr_dla_lock(mreg);
ARMCII_Buf_acc_scale(orig_bufs[i], new_bufs[i], size, datatype, scale);
if (mreg != NULL)
gmr_dla_unlock(mreg);
} else {
new_bufs[i] = orig_bufs[i];
}
if (mreg != NULL) {
// If the buffer wasn't copied, we should copy it into a private buffer
if (new_bufs[i] == orig_bufs[i]) {
MPI_Alloc_mem(size, MPI_INFO_NULL, &new_bufs[i]);
ARMCII_Assert(new_bufs[i] != NULL);
gmr_dla_lock(mreg);
ARMCI_Copy(orig_bufs[i], new_bufs[i], size);
gmr_dla_unlock(mreg);
}
}
if (new_bufs[i] == orig_bufs[i])
num_moved++;
}
*new_bufs_ptr = new_bufs;
return num_moved;
}
/** One-sided accumulate operation.
*
* @param[in] datatype ARMCI data type for the accumulate operation (see armci.h)
* @param[in] scale Pointer for a scalar of type datatype that will be used to
* scale values in the source buffer
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] bytes Number of bytes to transfer
* @param[in] proc Process id to target
* @return 0 on success, non-zero on failure
*/
int PARMCI_Acc(int datatype, void *scale, void *src, void *dst, int bytes, int proc) {
void *src_buf;
int count, type_size, scaled;
MPI_Datatype type;
gmr_t *src_mreg, *dst_mreg;
/* If NOGUARD is set, assume the buffer is not shared */
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
src_mreg = gmr_lookup(src, ARMCI_GROUP_WORLD.rank);
else
src_mreg = NULL;
dst_mreg = gmr_lookup(dst, proc);
ARMCII_Assert_msg(dst_mreg != NULL, "Invalid remote pointer");
/* Prepare the input data: Apply scaling if needed and acquire the DLA lock if
* needed. We hold the DLA lock if (src_buf == src && src_mreg != NULL). */
scaled = ARMCII_Buf_acc_is_scaled(datatype, scale);
if (scaled) {
MPI_Alloc_mem(bytes, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
ARMCII_Buf_acc_scale(src, src_buf, bytes, datatype, scale);
} else {
src_buf = src;
}
/* Check if we need to copy: user requested it or same mem region */
if ( (src_buf == src) /* buf_prepare didn't make a copy */
&& (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_COPY || src_mreg == dst_mreg) )
{
MPI_Alloc_mem(bytes, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
ARMCI_Copy(src, src_buf, bytes);
}
ARMCII_Acc_type_translate(datatype, &type, &type_size);
count = bytes/type_size;
ARMCII_Assert_msg(bytes % type_size == 0,
"Transfer size is not a multiple of the datatype size");
/* TODO: Support a local accumulate operation more efficiently */
gmr_accumulate(dst_mreg, src_buf, dst, count, type, proc);
gmr_flush(dst_mreg, proc, 1); /* flush_local */
if (src_buf != src)
MPI_Free_mem(src_buf);
return 0;
}
/** Blocking operation that accumulates data from the local process into the
* memory of the remote process. The data transfer is strided and blocking.
*
* @param[in] datatype Type of data to be transferred.
* @param[in] scale Pointer to the value that input data should be scaled by.
* @param[in] src_ptr Source starting address of the data block to put.
* @param[in] src_stride_arr Source array of stride distances in bytes.
* @param[in] dst_ptr Destination starting address to put data.
* @param[in] dst_stride_ar Destination array of stride distances in bytes.
* @param[in] count Block size in each dimension. count[0] should be the
* number of bytes of contiguous data in leading dimension.
* @param[in] stride_levels The level of strides.
* @param[in] proc Remote process ID (destination).
*
* @return Zero on success, error code otherwise.
*/
int PARMCI_AccS(int datatype, void *scale,
void *src_ptr, int src_stride_ar[/*stride_levels*/],
void *dst_ptr, int dst_stride_ar[/*stride_levels*/],
int count[/*stride_levels+1*/], int stride_levels, int proc) {
int err;
if (ARMCII_GLOBAL_STATE.strided_method == ARMCII_STRIDED_DIRECT) {
void *src_buf = NULL;
gmr_t *mreg, *gmr_loc = NULL;
MPI_Datatype src_type, dst_type, mpi_datatype;
int scaled, mpi_datatype_size;
ARMCII_Acc_type_translate(datatype, &mpi_datatype, &mpi_datatype_size);
scaled = ARMCII_Buf_acc_is_scaled(datatype, scale);
/* SCALE: copy and scale if requested */
if (scaled) {
armci_giov_t iov;
int i, nelem;
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
for (i = 1, nelem = count[0]/mpi_datatype_size; i < stride_levels+1; i++)
nelem *= count[i];
MPI_Alloc_mem(nelem*mpi_datatype_size, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
if (gmr_loc != NULL) gmr_dla_lock(gmr_loc);
/* Shoehorn the strided information into an IOV */
ARMCII_Strided_to_iov(&iov, src_ptr, src_stride_ar, src_ptr, src_stride_ar, count, stride_levels);
for (i = 0; i < iov.ptr_array_len; i++)
ARMCII_Buf_acc_scale(iov.src_ptr_array[i], ((uint8_t*)src_buf) + i*iov.bytes, iov.bytes, datatype, scale);
free(iov.src_ptr_array);
free(iov.dst_ptr_array);
if (gmr_loc != NULL) gmr_dla_unlock(gmr_loc);
MPI_Type_contiguous(nelem, mpi_datatype, &src_type);
}
/* COPY: Guard shared buffers */
else if (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_COPY) {
gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
if (gmr_loc != NULL) {
int i, nelem;
for (i = 1, nelem = count[0]/mpi_datatype_size; i < stride_levels+1; i++)
nelem *= count[i];
MPI_Alloc_mem(nelem*mpi_datatype_size, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
gmr_dla_lock(gmr_loc);
armci_write_strided(src_ptr, stride_levels, src_stride_ar, count, src_buf);
gmr_dla_unlock(gmr_loc);
MPI_Type_contiguous(nelem, mpi_datatype, &src_type);
}
}
/* NOGUARD: If src_buf hasn't been assigned to a copy, the strided source
* buffer is going to be used directly. */
if (src_buf == NULL) {
src_buf = src_ptr;
ARMCII_Strided_to_dtype(src_stride_ar, count, stride_levels, mpi_datatype, &src_type);
}
ARMCII_Strided_to_dtype(dst_stride_ar, count, stride_levels, mpi_datatype, &dst_type);
MPI_Type_commit(&src_type);
MPI_Type_commit(&dst_type);
int src_size, dst_size;
MPI_Type_size(src_type, &src_size);
MPI_Type_size(dst_type, &dst_size);
ARMCII_Assert(src_size == dst_size);
mreg = gmr_lookup(dst_ptr, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid shared pointer");
gmr_lock(mreg, proc);
gmr_accumulate_typed(mreg, src_buf, 1, src_type, dst_ptr, 1, dst_type, proc);
gmr_unlock(mreg, proc);
MPI_Type_free(&src_type);
MPI_Type_free(&dst_type);
/* COPY/SCALE: Free temp buffer */
if (src_buf != src_ptr)
MPI_Free_mem(src_buf);
err = 0;
} else {
armci_giov_t iov;
ARMCII_Strided_to_iov(&iov, src_ptr, src_stride_ar, dst_ptr, dst_stride_ar, count, stride_levels);
err = PARMCI_AccV(datatype, scale, &iov, 1, proc);
free(iov.src_ptr_array);
free(iov.dst_ptr_array);
}
return err;
}
