/** 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 put operation.
*
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] size Number of bytes to transfer
* @param[in] target Process id to target
* @return 0 on success, non-zero on failure
*/
int ARMCI_Put(void *src, void *dst, int size, int target) {
gmr_t *src_mreg, *dst_mreg;
src_mreg = gmr_lookup(src, ARMCI_GROUP_WORLD.rank);
dst_mreg = gmr_lookup(dst, target);
ARMCII_Assert_msg(dst_mreg != NULL, "Invalid remote pointer");
/* Local operation */
if (target == ARMCI_GROUP_WORLD.rank) {
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
gmr_dla_lock(dst_mreg);
if (src_mreg) gmr_dla_lock(src_mreg);
}
ARMCI_Copy(src, dst, size);
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
gmr_dla_unlock(dst_mreg);
if (src_mreg) gmr_dla_unlock(src_mreg);
}
}
/* Origin buffer is private */
else if (src_mreg == NULL || ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_NOGUARD) {
gmr_lock(dst_mreg, target);
gmr_put(dst_mreg, src, dst, size, target);
gmr_unlock(dst_mreg, target);
}
/* COPY: Either origin and target buffers are in the same window and we can't
* lock the same window twice (MPI semantics) or the user has requested
* always-copy mode. */
else {
void *src_buf;
MPI_Alloc_mem(size, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
gmr_dla_lock(src_mreg);
ARMCI_Copy(src, src_buf, size);
gmr_dla_unlock(src_mreg);
gmr_lock(dst_mreg, target);
gmr_put(dst_mreg, src_buf, dst, size, target);
gmr_unlock(dst_mreg, target);
MPI_Free_mem(src_buf);
}
return 0;
}
/** Declare the end of a local access epoch.
*
* \note MPI-2 does not allow multiple locks at once, so you can have only one
* access epoch open at a time and cannot do put/get/acc while in an access
* region.
*
* @param[in] ptr Pointer to the allocation that was accessed directly
*/
void ARMCI_Access_end(void *ptr) {
gmr_t *mreg;
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
gmr_dla_unlock(mreg);
}
/** Finish a set of prepared buffers. Will perform communication and copies as
* needed to ensure results are in the original buffers. Temporary space will be
* freed.
*
* @param[in] orig_bufs Original set of buffers.
* @param[out] new_bufs 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).
*/
void ARMCII_Buf_finish_write_vec(void **orig_bufs, void **new_bufs, int count, int size) {
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
int i;
for (i = 0; i < count; i++) {
if (orig_bufs[i] != new_bufs[i]) {
gmr_t *mreg = gmr_lookup(orig_bufs[i], ARMCI_GROUP_WORLD.rank);
ARMCII_Assert(mreg != NULL);
gmr_dla_lock(mreg);
ARMCI_Copy(new_bufs[i], orig_bufs[i], size);
// gmr_put(mreg, new_bufs[i], orig_bufs[i], size, ARMCI_GROUP_WORLD.rank);
gmr_dla_unlock(mreg);
MPI_Free_mem(new_bufs[i]);
}
}
free(new_bufs);
}
}
/** Prepare a set of buffers for use with a put operation. The returned set of
* buffers is guaranteed to be in private space. 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).
* @return Number of buffers that were moved.
*/
int ARMCII_Buf_prepare_read_vec(void **orig_bufs, void ***new_bufs_ptr, int count, int size) {
int num_moved = 0;
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
void **new_bufs = malloc(count*sizeof(void*));
int i;
for (i = 0; i < count; i++)
new_bufs[i] = NULL;
for (i = 0; i < count; i++) {
// Check if the source buffer is within a shared region. If so, copy it
// into a private buffer.
gmr_t *mreg = gmr_lookup(orig_bufs[i], ARMCI_GROUP_WORLD.rank);
if (mreg != NULL) {
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_get(mreg, orig_bufs[i], new_bufs[i], size, ARMCI_GROUP_WORLD.rank);
gmr_dla_unlock(mreg);
num_moved++;
} else {
new_bufs[i] = orig_bufs[i];
}
}
*new_bufs_ptr = new_bufs;
}
else {
*new_bufs_ptr = orig_bufs;
}
return num_moved;
}
/** Destroy/free a shared memory region.
*
* @param[in] ptr Pointer within range of the segment (e.g. base pointer).
* @param[in] group Group on which to perform the free.
*/
void gmr_destroy(gmr_t *mreg, ARMCI_Group *group) {
int search_proc_in, search_proc_out, search_proc_out_grp;
void *search_base;
int alloc_me, alloc_nproc;
int world_me, world_nproc;
MPI_Comm_rank(group->comm, &alloc_me);
MPI_Comm_size(group->comm, &alloc_nproc);
MPI_Comm_rank(ARMCI_GROUP_WORLD.comm, &world_me);
MPI_Comm_size(ARMCI_GROUP_WORLD.comm, &world_nproc);
/* All-to-all exchange of a <base address, proc> pair. This is so that we
* can support passing NULL into ARMCI_Free() which is permitted when a
* process allocates 0 bytes. Unfortunately, in this case we still need to
* identify the mem region and free it.
*/
if (mreg == NULL)
search_proc_in = -1;
else {
search_proc_in = world_me;
search_base = mreg->slices[world_me].base;
}
/* Collectively decide on who will provide the base address */
MPI_Allreduce(&search_proc_in, &search_proc_out, 1, MPI_INT, MPI_MAX, group->comm);
/* Everyone passed NULL. Nothing to free. */
if (search_proc_out < 0)
return;
/* Translate world rank to group rank */
search_proc_out_grp = ARMCII_Translate_absolute_to_group(group, search_proc_out);
/* Broadcast the base address */
MPI_Bcast(&search_base, sizeof(void*), MPI_BYTE, search_proc_out_grp, group->comm);
/* If we were passed NULL, look up the mem region using the <base, proc> pair */
if (mreg == NULL)
mreg = gmr_lookup(search_base, search_proc_out);
/* If it's still not found, the user may have passed the wrong group */
ARMCII_Assert_msg(mreg != NULL, "Could not locate the desired allocation");
switch (mreg->lock_state) {
case GMR_LOCK_UNLOCKED:
break;
case GMR_LOCK_DLA:
ARMCII_Warning("Releasing direct local access before freeing shared allocation\n");
gmr_dla_unlock(mreg);
break;
default:
ARMCII_Error("Unable to free locked memory region (%d)\n", mreg->lock_state);
}
/* Remove from the list of mem regions */
if (mreg->prev == NULL) {
ARMCII_Assert(gmr_list == mreg);
gmr_list = mreg->next;
if (mreg->next != NULL)
mreg->next->prev = NULL;
} else {
mreg->prev->next = mreg->next;
if (mreg->next != NULL)
mreg->next->prev = mreg->prev;
}
/* Destroy the window and free all buffers */
MPI_Win_free(&mreg->window);
if (mreg->slices[world_me].base != NULL)
MPI_Free_mem(mreg->slices[world_me].base);
free(mreg->slices);
ARMCIX_Destroy_mutexes_hdl(mreg->rmw_mutex);
free(mreg);
}
/** 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;
}
/** Blocking operation that transfers data from the remote process to the
* memory of the calling process. The data transfer is strided and blocking.
*
* @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_GetS(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 *dst_buf = NULL;
gmr_t *mreg, *gmr_loc = NULL;
MPI_Datatype src_type, dst_type;
/* COPY: Guard shared buffers */
if (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_COPY) {
gmr_loc = gmr_lookup(dst_ptr, ARMCI_GROUP_WORLD.rank);
if (gmr_loc != NULL) {
int i, size;
for (i = 1, size = count[0]; i < stride_levels+1; i++)
size *= count[i];
MPI_Alloc_mem(size, MPI_INFO_NULL, &dst_buf);
ARMCII_Assert(dst_buf != NULL);
MPI_Type_contiguous(size, MPI_BYTE, &dst_type);
}
}
/* NOGUARD: If dst_buf hasn't been assigned to a copy, the strided source
* buffer is going to be used directly. */
if (dst_buf == NULL) {
dst_buf = dst_ptr;
ARMCII_Strided_to_dtype(dst_stride_ar, count, stride_levels, MPI_BYTE, &dst_type);
}
ARMCII_Strided_to_dtype(src_stride_ar, count, stride_levels, MPI_BYTE, &src_type);
MPI_Type_commit(&src_type);
MPI_Type_commit(&dst_type);
mreg = gmr_lookup(src_ptr, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid shared pointer");
gmr_lock(mreg, proc);
gmr_get_typed(mreg, src_ptr, 1, src_type, dst_buf, 1, dst_type, proc);
gmr_unlock(mreg, proc);
/* COPY: Finish the transfer */
if (dst_buf != dst_ptr) {
gmr_dla_lock(gmr_loc);
armci_read_strided(dst_ptr, stride_levels, dst_stride_ar, count, dst_buf);
gmr_dla_unlock(gmr_loc);
MPI_Free_mem(dst_buf);
}
MPI_Type_free(&src_type);
MPI_Type_free(&dst_type);
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_GetV(&iov, 1, proc);
free(iov.src_ptr_array);
free(iov.dst_ptr_array);
}
return err;
}
/** 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 ARMCI_Acc(int datatype, void *scale, void *src, void *dst, int bytes, int proc) {
void *src_buf;
int count, type_size, scaled, src_is_locked = 0;
MPI_Datatype type;
gmr_t *src_mreg, *dst_mreg;
src_mreg = gmr_lookup(src, ARMCI_GROUP_WORLD.rank);
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 (src_mreg && ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
gmr_dla_lock(src_mreg);
src_is_locked = 1;
}
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);
}
/* Unlock early if src_buf is a copy */
if (src_buf != src && src_is_locked) {
gmr_dla_unlock(src_mreg);
src_is_locked = 0;
}
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_lock(dst_mreg, proc);
gmr_accumulate(dst_mreg, src_buf, dst, count, type, proc);
gmr_unlock(dst_mreg, proc);
if (src_is_locked) {
gmr_dla_unlock(src_mreg);
src_is_locked = 0;
}
if (src_buf != src)
MPI_Free_mem(src_buf);
return 0;
}