/** One-sided accumulate operation with typed arguments. Source buffer must be private.
*
* @param[in] mreg Memory region
* @param[in] src Address of source data
* @param[in] src_count Number of elements of the given type at the source
* @param[in] src_type MPI datatype of the source elements
* @param[in] dst Address of destination buffer
* @param[in] dst_count Number of elements of the given type at the destination
* @param[in] src_type MPI datatype of the destination elements
* @param[in] size Number of bytes to transfer
* @param[in] proc Absolute process id of target process
* @return 0 on success, non-zero on failure
*/
int gmr_accumulate_typed(gmr_t *mreg, void *src, int src_count, MPI_Datatype src_type,
void *dst, int dst_count, MPI_Datatype dst_type, int proc) {
int grp_proc;
gmr_size_t disp;
MPI_Aint lb, extent;
grp_proc = ARMCII_Translate_absolute_to_group(&mreg->group, proc);
ARMCII_Assert(grp_proc >= 0);
// Calculate displacement from beginning of the window
if (dst == MPI_BOTTOM)
disp = 0;
else
disp = (gmr_size_t) ((uint8_t*)dst - (uint8_t*)mreg->slices[proc].base);
// Perform checks
MPI_Type_get_true_extent(dst_type, &lb, &extent);
ARMCII_Assert(mreg->lock_state != GMR_LOCK_UNLOCKED);
ARMCII_Assert_msg(disp >= 0 && disp < mreg->slices[proc].size, "Invalid remote address");
ARMCII_Assert_msg(disp + dst_count*extent <= mreg->slices[proc].size, "Transfer is out of range");
MPI_Accumulate(src, src_count, src_type, grp_proc, (MPI_Aint) disp, dst_count, dst_type, MPI_SUM, mreg->window);
return 0;
}
/** 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;
}
/** Declare the start of a local access epoch. This allows direct access to
* data in local memory.
*
* @param[in] ptr Pointer to the allocation that will be accessed directly
*/
void ARMCI_Access_begin(void *ptr) {
gmr_t *mreg;
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
ARMCII_Assert_msg((mreg->access_mode & ARMCIX_MODE_NO_LOAD_STORE) == 0,
"Direct local access is not permitted in the current access mode");
gmr_dla_lock(mreg);
}
/** Perform an I/O vector operation. Local buffers must be private.
*
* @param[in] op Operation to be performed (ARMCII_OP_PUT, ...)
* @param[in] src Array of source pointers
* @param[in] dst Array of destination pointers
* @param[in] count Length of pointer arrays
* @param[in] size Size of each transfer
* @param[in] datatype Data type for accumulate op (ignored for all others)
* @param[in] overlapping Do remote regions overlap?
* @param[in] same_alloc Do remote regions correspond to the same allocation?
* @param[in] proc Target process
* @return Zero on success, error code otherwise
*/
int ARMCII_Iov_op_dispatch(enum ARMCII_Op_e op, void **src, void **dst, int count, int size,
int datatype, int overlapping, int same_alloc, int proc) {
MPI_Datatype type;
int type_count, type_size;
if (op == ARMCII_OP_ACC) {
ARMCII_Acc_type_translate(datatype, &type, &type_size);
type_count = size/type_size;
ARMCII_Assert_msg(size % type_size == 0, "Transfer size is not a multiple of type size");
} else {
type = MPI_BYTE;
MPI_Type_size(type, &type_size);
type_count = size/type_size;
ARMCII_Assert_msg(size % type_size == 0, "Transfer size is not a multiple of type size");
}
// CONSERVATIVE CASE: If remote pointers overlap or remote pointers correspond to
// multiple allocations, use the safe implementation to avoid invalid MPI
// use.
if (overlapping || !same_alloc || ARMCII_GLOBAL_STATE.iov_method == ARMCII_IOV_CONSRV) {
if (overlapping) ARMCII_Warning("IOV remote buffers overlap\n");
if (!same_alloc) ARMCII_Warning("IOV remote buffers are not within the same allocation\n");
return ARMCII_Iov_op_safe(op, src, dst, count, type_count, type, proc);
}
// OPTIMIZED CASE: It's safe for us to issue all the operations under a
// single lock.
else if ( ARMCII_GLOBAL_STATE.iov_method == ARMCII_IOV_DIRECT
|| ARMCII_GLOBAL_STATE.iov_method == ARMCII_IOV_AUTO ) {
if (ARMCII_GLOBAL_STATE.no_mpi_bottom == 1) {
return ARMCII_Iov_op_datatype_no_bottom(op, src, dst, count, type_count, type, proc);
} else {
return ARMCII_Iov_op_datatype(op, src, dst, count, type_count, type, proc);
}
} else if (ARMCII_GLOBAL_STATE.iov_method == ARMCII_IOV_BATCHED) {
return ARMCII_Iov_op_batched(op, src, dst, count, type_count, type, proc);
} else {
ARMCII_Error("unknown iov method (%d)\n", ARMCII_GLOBAL_STATE.iov_method);
return 1;
}
}
/** Query the access mode for the given allocation. Non-collective.
*
* @param[in] ptr Pointer within the allocation.
* @return Current access mode.
*/
int ARMCIX_Mode_get(void *ptr) {
gmr_t *mreg;
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
return mreg->access_mode;
}
/** 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 PARMCI_Access_end(void *ptr) {
gmr_t *mreg;
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
gmr_sync(mreg);
}
/** Optimized implementation of the ARMCI IOV operation that uses a single
* lock/unlock pair.
*/
int ARMCII_Iov_op_batched(enum ARMCII_Op_e op, void **src, void **dst, int count, int elem_count,
MPI_Datatype type, int proc) {
int i;
gmr_t *mreg;
void *shr_ptr;
switch(op) {
case ARMCII_OP_PUT:
shr_ptr = dst[0];
break;
case ARMCII_OP_GET:
shr_ptr = src[0];
break;
case ARMCII_OP_ACC:
shr_ptr = dst[0];
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
mreg = gmr_lookup(shr_ptr, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
gmr_lock(mreg, proc);
for (i = 0; i < count; i++) {
if ( ARMCII_GLOBAL_STATE.iov_batched_limit > 0
&& i % ARMCII_GLOBAL_STATE.iov_batched_limit == 0
&& i > 0 )
{
gmr_unlock(mreg, proc);
gmr_lock(mreg, proc);
}
switch(op) {
case ARMCII_OP_PUT:
gmr_put(mreg, src[i], dst[i], elem_count, proc);
break;
case ARMCII_OP_GET:
gmr_get(mreg, src[i], dst[i], elem_count, proc);
break;
case ARMCII_OP_ACC:
gmr_accumulate(mreg, src[i], dst[i], elem_count, type, proc);
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
}
gmr_unlock(mreg, proc);
return 0;
}
/** Perform atomic read-modify-write on the given integer or long location and
* return the location's original value.
*
* \note ARMCI RMW operations are atomic with respect to other RMW operations,
* but not with respect to other one-sided operations (get, put, acc, etc).
*
* @param[in] op Operation to be performed:
* ARMCI_FETCH_AND_ADD (int)
* ARMCI_FETCH_AND_ADD_LONG
* ARMCI_SWAP (int)
* ARMCI_SWAP_LONG
* @param[out] ploc Location to store the original value.
* @param[in] prem Location on which to perform atomic operation.
* @param[in] value Value to add to remote location (ignored for swap).
* @param[in] proc Process rank for the target buffer.
*/
int PARMCI_Rmw(int op, void *ploc, void *prem, int value, int proc) {
int is_long;
gmr_t *mreg;
mreg = gmr_lookup(prem, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
if (op == ARMCI_SWAP_LONG || op == ARMCI_FETCH_AND_ADD_LONG)
is_long = 1;
else
is_long = 0;
if (op == ARMCI_SWAP || op == ARMCI_SWAP_LONG) {
long swap_val_l;
int swap_val_i;
ARMCIX_Lock_hdl(mreg->rmw_mutex, 0, proc);
PARMCI_Get(prem, is_long ? (void*) &swap_val_l : (void*) &swap_val_i,
is_long ? sizeof(long) : sizeof(int), proc);
PARMCI_Put(ploc, prem, is_long ? sizeof(long) : sizeof(int), proc);
ARMCIX_Unlock_hdl(mreg->rmw_mutex, 0, proc);
if (is_long)
*(long*) ploc = swap_val_l;
else
*(int*) ploc = swap_val_i;
}
else if (op == ARMCI_FETCH_AND_ADD || op == ARMCI_FETCH_AND_ADD_LONG) {
long fetch_val_l, new_val_l;
int fetch_val_i, new_val_i;
ARMCIX_Lock_hdl(mreg->rmw_mutex, 0, proc);
PARMCI_Get(prem, is_long ? (void*) &fetch_val_l : (void*) &fetch_val_i,
is_long ? sizeof(long) : sizeof(int), proc);
if (is_long)
new_val_l = fetch_val_l + value;
else
new_val_i = fetch_val_i + value;
PARMCI_Put(is_long ? (void*) &new_val_l : (void*) &new_val_i, prem,
is_long ? sizeof(long) : sizeof(int), proc);
ARMCIX_Unlock_hdl(mreg->rmw_mutex, 0, proc);
if (is_long)
*(long*) ploc = fetch_val_l;
else
*(int*) ploc = fetch_val_i;
}
else {
ARMCII_Error("invalid operation (%d)", op);
}
return 0;
}
/** Convert an ARMCI strided access description into an MPI subarray datatype.
*
* @param[in] stride_array Array of strides
* @param[in] count Array of transfer counts
* @param[in] stride_levels Number of levels of striding
* @param[in] old_type Type of the data element described by count and stride_array
* @param[out] new_type New MPI type for the given strided access
*/
void ARMCII_Strided_to_dtype(int stride_array[/*stride_levels*/], int count[/*stride_levels+1*/],
int stride_levels, MPI_Datatype old_type, MPI_Datatype *new_type)
{
int sizes [stride_levels+1];
int subsizes[stride_levels+1];
int starts [stride_levels+1];
int i, old_type_size;
MPI_Type_size(old_type, &old_type_size);
/* Eliminate counts that don't count (all 1 counts at the end) */
for (i = stride_levels+1; i > 0 && stride_levels > 0 && count[i-1] == 1; i--)
stride_levels--;
/* A correct strided spec should me monotonic increasing and stride_array[i+1] should
be a multiple of stride_array[i]. */
if (stride_levels > 0) {
for (i = 1; i < stride_levels; i++)
ARMCII_Assert(stride_array[i] >= stride_array[i-1] && stride_array[i] % stride_array[i-1] == 0);
}
/* Test for a contiguous transfer */
if (stride_levels == 0) {
int elem_count = count[0]/old_type_size;
ARMCII_Assert(count[0] % old_type_size == 0);
MPI_Type_contiguous(elem_count, old_type, new_type);
}
/* Transfer is non-contiguous */
else {
for (i = 0; i < stride_levels+1; i++)
starts[i] = 0;
sizes [stride_levels] = stride_array[0]/old_type_size;
subsizes[stride_levels] = count[0]/old_type_size;
ARMCII_Assert(stride_array[0] % old_type_size == 0 && count[0] % old_type_size == 0);
for (i = 1; i < stride_levels; i++) {
/* Convert strides into dimensions by dividing out contributions from lower dims */
sizes [stride_levels-i] = stride_array[i]/stride_array[i-1];
subsizes[stride_levels-i] = count[i];
ARMCII_Assert_msg(stride_array[i] % stride_array[i-1] == 0, "Invalid striding");
}
sizes [0] = count[stride_levels];
subsizes[0] = count[stride_levels];
MPI_Type_create_subarray(stride_levels+1, sizes, subsizes, starts, MPI_ORDER_C, old_type, new_type);
}
}
/** Set the acess mode for the given allocation. Collective across the
* allocation's group. Waits for all processes, finishes all communication,
* and then sets the new access mode.
*
* @param[in] new_mode The new access mode.
* @param[in] ptr Pointer within the allocation.
* @return Zero upon success, error code otherwise.
*/
int ARMCIX_Mode_set(int new_mode, void *ptr, ARMCI_Group *group) {
gmr_t *mreg;
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
ARMCII_Assert(group->comm == mreg->group.comm);
ARMCII_Assert_msg(mreg->lock_state != GMR_LOCK_DLA,
"Cannot change the access mode; window is locked for local access.");
ARMCII_Assert_msg(mreg->lock_state == GMR_LOCK_UNLOCKED,
"Cannot change the access mode on a window that is locked.");
// Wait for all processes to complete any outstanding communication before we
// do the mode switch
MPI_Barrier(mreg->group.comm);
mreg->access_mode = new_mode;
return 0;
}
/** 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;
}
/** Free a shared memory allocation. Collective.
*
* @param[in] ptr Pointer to the local patch of the allocation
*/
int ARMCI_Free_group(void *ptr, ARMCI_Group *group) {
gmr_t *mreg;
if (ptr != NULL) {
mreg = gmr_lookup(ptr, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(mreg != NULL, "Invalid shared pointer");
} else {
ARMCII_Dbg_print(DEBUG_CAT_ALLOC, "given NULL\n");
mreg = NULL;
}
gmr_destroy(mreg, group);
return 0;
}
/** Unlock a memory region that was locked for direct local access.
*
* @param[in] mreg Memory region
*/
void gmr_dla_unlock(gmr_t *mreg) {
int grp_proc = ARMCII_Translate_absolute_to_group(&mreg->group, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert(grp_proc >= 0);
ARMCII_Assert(mreg->lock_state == GMR_LOCK_DLA);
ARMCII_Assert_msg((mreg->access_mode & ARMCIX_MODE_NO_LOAD_STORE) == 0,
"Direct local access is not allowed in the current access mode");
mreg->dla_lock_count--;
if (mreg->dla_lock_count == 0) {
MPI_Win_unlock(grp_proc, mreg->window);
mreg->lock_state = GMR_LOCK_UNLOCKED;
}
}
/** Create a group of ARMCI mutexes. Collective onthe ARMCI group.
*
* @param[in] count Number of mutexes on the local process.
* @param[in] pgroup ARMCI group on which to create mutexes
* @return Handle to the mutex group.
*/
armcix_mutex_hdl_t ARMCIX_Create_mutexes_hdl(int my_count, ARMCI_Group *pgroup) {
int rank, nproc, max_count, i;
armcix_mutex_hdl_t hdl;
hdl = malloc(sizeof(struct armcix_mutex_hdl_s));
ARMCII_Assert(hdl != NULL);
ARMCIX_Group_dup(pgroup, &hdl->grp);
MPI_Comm_rank(hdl->grp.comm, &rank);
MPI_Comm_size(hdl->grp.comm, &nproc);
hdl->my_count = my_count;
/* Find the max. count to determine how many windows we need. */
MPI_Allreduce(&my_count, &max_count, 1, MPI_INT, MPI_MAX, hdl->grp.comm);
ARMCII_Assert_msg(max_count > 0, "Invalid number of mutexes");
hdl->max_count = max_count;
hdl->windows = malloc(sizeof(MPI_Win)*max_count);
if (my_count > 0) {
hdl->bases = malloc(sizeof(uint8_t*)*my_count);
} else {
hdl->bases = NULL;
}
/* We need multiple windows here: one for each mutex. Otherwise
performance will suffer due to exclusive access epochs. */
for (i = 0; i < max_count; i++) {
int size = 0;
void *base = NULL;
if (i < my_count) {
MPI_Alloc_mem(nproc, MPI_INFO_NULL, &hdl->bases[i]);
ARMCII_Assert(hdl->bases[i] != NULL);
ARMCII_Bzero(hdl->bases[i], nproc);
base = hdl->bases[i];
size = nproc;
}
MPI_Win_create(base, size, sizeof(uint8_t), MPI_INFO_NULL, hdl->grp.comm, &hdl->windows[i]);
}
return hdl;
}
/** One-sided get 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 PARMCI_Get(void *src, void *dst, int size, int target) {
gmr_t *src_mreg, *dst_mreg;
src_mreg = gmr_lookup(src, target);
/* If NOGUARD is set, assume the buffer is not shared */
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
dst_mreg = gmr_lookup(dst, ARMCI_GROUP_WORLD.rank);
else
dst_mreg = NULL;
ARMCII_Assert_msg(src_mreg != NULL, "Invalid remote pointer");
/* Local operation */
if (target == ARMCI_GROUP_WORLD.rank && dst_mreg == NULL) {
ARMCI_Copy(src, dst, size);
}
/* Origin buffer is private */
else if (dst_mreg == NULL) {
gmr_get(src_mreg, src, dst, size, target);
gmr_flush(src_mreg, target, 0); /* it's a round trip so w.r.t. flush, local=remote */
}
/* 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 *dst_buf;
MPI_Alloc_mem(size, MPI_INFO_NULL, &dst_buf);
ARMCII_Assert(dst_buf != NULL);
gmr_get(src_mreg, src, dst_buf, size, target);
gmr_flush(src_mreg, target, 0); /* it's a round trip so w.r.t. flush, local=remote */
ARMCI_Copy(dst_buf, dst, size);
MPI_Free_mem(dst_buf);
}
return 0;
}
/** One-sided accumulate operation. Source buffer must be private.
*
* @param[in] mreg Memory region
* @param[in] src Source address (local)
* @param[in] dst Destination address (remote)
* @param[in] type MPI type of the given buffers
* @param[in] count Number of elements of the given type to transfer
* @param[in] proc Absolute process id of the target
* @return 0 on success, non-zero on failure
*/
int gmr_accumulate(gmr_t *mreg, void *src, void *dst, int count, MPI_Datatype type, int proc) {
ARMCII_Assert_msg(src != NULL, "Invalid local address");
return gmr_accumulate_typed(mreg, src, count, type, dst, count, type, proc);
}
/** One-sided get operation. Destination buffer must be private.
*
* @param[in] mreg Memory region
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] size Number of bytes to transfer
* @param[in] proc Absolute process id of target process
* @return 0 on success, non-zero on failure
*/
int gmr_get(gmr_t *mreg, void *src, void *dst, int size, int proc) {
ARMCII_Assert_msg(dst != NULL, "Invalid local address");
return gmr_get_typed(mreg, src, size, MPI_BYTE, dst, size, MPI_BYTE, proc);
}
/** 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);
}
/** 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] buf Original set of 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 Pointer to the new buffer or buf
*/
void ARMCII_Buf_acc_scale(void *buf_in, void *buf_out, int size, int datatype, void *scale) {
int j, nelem;
int type_size = -1;
MPI_Datatype type;
switch (datatype) {
case ARMCI_ACC_INT:
MPI_Type_size(MPI_INT, &type_size);
type = MPI_INT;
nelem= size/type_size;
{
int *src_i = (int*) buf_in;
int *scl_i = (int*) buf_out;
const int s = *((int*) scale);
for (j = 0; j < nelem; j++)
scl_i[j] = src_i[j]*s;
}
break;
case ARMCI_ACC_LNG:
MPI_Type_size(MPI_LONG, &type_size);
type = MPI_LONG;
nelem= size/type_size;
{
long *src_l = (long*) buf_in;
long *scl_l = (long*) buf_out;
const long s = *((long*) scale);
for (j = 0; j < nelem; j++)
scl_l[j] = src_l[j]*s;
}
break;
case ARMCI_ACC_FLT:
MPI_Type_size(MPI_FLOAT, &type_size);
type = MPI_FLOAT;
nelem= size/type_size;
{
float *src_f = (float*) buf_in;
float *scl_f = (float*) buf_out;
const float s = *((float*) scale);
for (j = 0; j < nelem; j++)
scl_f[j] = src_f[j]*s;
}
break;
case ARMCI_ACC_DBL:
MPI_Type_size(MPI_DOUBLE, &type_size);
type = MPI_DOUBLE;
nelem= size/type_size;
{
double *src_d = (double*) buf_in;
double *scl_d = (double*) buf_out;
const double s = *((double*) scale);
for (j = 0; j < nelem; j++)
scl_d[j] = src_d[j]*s;
}
break;
case ARMCI_ACC_CPL:
MPI_Type_size(MPI_FLOAT, &type_size);
type = MPI_FLOAT;
nelem= size/type_size;
{
float *src_fc = (float*) buf_in;
float *scl_fc = (float*) buf_out;
const float s_r = ((float*)scale)[0];
const float s_c = ((float*)scale)[1];
for (j = 0; j < nelem; j += 2) {
// Complex multiplication: (a + bi)*(c + di)
const float src_fc_j = src_fc[j];
const float src_fc_j_1 = src_fc[j+1];
/*
scl_fc[j] = src_fc[j]*s_r - src_fc[j+1]*s_c;
scl_fc[j+1] = src_fc[j+1]*s_r + src_fc[j]*s_c;
*/
scl_fc[j] = src_fc_j*s_r - src_fc_j_1*s_c;
scl_fc[j+1] = src_fc_j_1*s_r + src_fc_j*s_c;
}
}
break;
case ARMCI_ACC_DCP:
MPI_Type_size(MPI_DOUBLE, &type_size);
type = MPI_DOUBLE;
nelem= size/type_size;
{
double *src_dc = (double*) buf_in;
double *scl_dc = (double*) buf_out;
const double s_r = ((double*)scale)[0];
const double s_c = ((double*)scale)[1];
for (j = 0; j < nelem; j += 2) {
// Complex multiplication: (a + bi)*(c + di)
const double src_dc_j = src_dc[j];
const double src_dc_j_1 = src_dc[j+1];
/*
scl_dc[j] = src_dc[j]*s_r - src_dc[j+1]*s_c;
scl_dc[j+1] = src_dc[j+1]*s_r + src_dc[j]*s_c;
*/
scl_dc[j] = src_dc_j*s_r - src_dc_j_1*s_c;
scl_dc[j+1] = src_dc_j_1*s_r + src_dc_j*s_c;
}
}
break;
default:
ARMCII_Error("unknown data type (%d)", datatype);
}
ARMCII_Assert_msg(size % type_size == 0,
"Transfer size is not a multiple of the datatype size");
}
/** 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;
}
/** Optimized implementation of the ARMCI IOV operation that uses an MPI
* datatype to achieve a one-sided gather/scatter. Does not use MPI_BOTTOM.
*/
int ARMCII_Iov_op_datatype_no_bottom(enum ARMCII_Op_e op, void **src, void **dst, int count, int elem_count,
MPI_Datatype type, int proc) {
gmr_t *mreg;
MPI_Datatype type_loc, type_rem;
MPI_Aint disp_loc[count];
int disp_rem[count];
int block_len[count];
void *dst_win_base;
int dst_win_size, i, type_size;
void **buf_rem, **buf_loc;
MPI_Aint base_rem;
MPI_Aint base_loc;
void *base_loc_ptr;
switch(op) {
case ARMCII_OP_ACC:
case ARMCII_OP_PUT:
buf_rem = dst;
buf_loc = src;
break;
case ARMCII_OP_GET:
buf_rem = src;
buf_loc = dst;
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
MPI_Type_size(type, &type_size);
mreg = gmr_lookup(buf_rem[0], proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
dst_win_base = mreg->slices[proc].base;
dst_win_size = mreg->slices[proc].size;
MPI_Get_address(dst_win_base, &base_rem);
/* Pick a base address for the start of the origin's datatype */
base_loc_ptr = buf_loc[0];
MPI_Get_address(base_loc_ptr, &base_loc);
for (i = 0; i < count; i++) {
MPI_Aint target_rem, target_loc;
MPI_Get_address(buf_loc[i], &target_loc);
MPI_Get_address(buf_rem[i], &target_rem);
disp_loc[i] = target_loc - base_loc;
disp_rem[i] = (target_rem - base_rem)/type_size;
block_len[i] = elem_count;
ARMCII_Assert_msg((target_rem - base_rem) % type_size == 0, "Transfer size is not a multiple of type size");
ARMCII_Assert_msg(disp_rem[i] >= 0 && disp_rem[i] < dst_win_size, "Invalid remote pointer");
ARMCII_Assert_msg(((uint8_t*)buf_rem[i]) + block_len[i] <= ((uint8_t*)dst_win_base) + dst_win_size, "Transfer exceeds buffer length");
}
MPI_Type_create_hindexed(count, block_len, disp_loc, type, &type_loc);
MPI_Type_create_indexed_block(count, elem_count, disp_rem, type, &type_rem);
//MPI_Type_indexed(count, block_len, disp_rem, type, &type_rem);
MPI_Type_commit(&type_loc);
MPI_Type_commit(&type_rem);
gmr_lock(mreg, proc);
switch(op) {
case ARMCII_OP_ACC:
gmr_accumulate_typed(mreg, base_loc_ptr, 1, type_loc, MPI_BOTTOM, 1, type_rem, proc);
break;
case ARMCII_OP_PUT:
gmr_put_typed(mreg, base_loc_ptr, 1, type_loc, MPI_BOTTOM, 1, type_rem, proc);
break;
case ARMCII_OP_GET:
gmr_get_typed(mreg, MPI_BOTTOM, 1, type_rem, base_loc_ptr, 1, type_loc, proc);
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
gmr_unlock(mreg, proc);
MPI_Type_free(&type_loc);
MPI_Type_free(&type_rem);
return 0;
}
/** 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;
}
/** Perform atomic read-modify-write on the given integer or long location and
* return the location's original value.
*
* \note ARMCI RMW operations are atomic with respect to other RMW operations,
* but not with respect to other one-sided operations (get, put, acc, etc).
*
* @param[in] op Operation to be performed:
* ARMCI_FETCH_AND_ADD (int)
* ARMCI_FETCH_AND_ADD_LONG
* ARMCI_SWAP (int)
* ARMCI_SWAP_LONG
* @param[out] ploc Location to store the original value.
* @param[in] prem Location on which to perform atomic operation.
* @param[in] value Value to add to remote location (ignored for swap).
* @param[in] proc Process rank for the target buffer.
*/
int PARMCI_Rmw(int op, void *ploc, void *prem, int value, int proc) {
int is_swap = 0, is_long = 0;
MPI_Datatype type;
MPI_Op rop;
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(ploc, ARMCI_GROUP_WORLD.rank);
else
src_mreg = NULL;
dst_mreg = gmr_lookup(prem, proc);
ARMCII_Assert_msg(dst_mreg != NULL, "Invalid remote pointer");
if (op == ARMCI_SWAP_LONG || op == ARMCI_FETCH_AND_ADD_LONG) {
is_long = 1;
type = MPI_LONG;
}
else
type = MPI_INT;
if (op == ARMCI_SWAP || op == ARMCI_SWAP_LONG) {
is_swap = 1;
rop = MPI_REPLACE;
}
else if (op == ARMCI_FETCH_AND_ADD || op == ARMCI_FETCH_AND_ADD_LONG)
rop = MPI_SUM;
else
ARMCII_Error("invalid operation (%d)", op);
/* We hold the DLA lock if (src_mreg != NULL). */
if (is_swap) {
long out_val_l, src_val_l = *((long*)ploc);
int out_val_i, src_val_i = *((int*)ploc);
gmr_fetch_and_op(dst_mreg,
is_long ? (void*) &src_val_l : (void*) &src_val_i /* src */,
is_long ? (void*) &out_val_l : (void*) &out_val_i /* out */,
prem /* dst */, type, rop, proc);
gmr_flush(dst_mreg, proc, 0); /* it's a round trip so w.r.t. flush, local=remote */
if (is_long)
*(long*) ploc = out_val_l;
else
*(int*) ploc = out_val_i;
}
else /* fetch-and-add */ {
long fetch_val_l, add_val_l = value;
int fetch_val_i, add_val_i = value;
gmr_fetch_and_op(dst_mreg,
is_long ? (void*) &add_val_l : (void*) &add_val_i /* src */,
is_long ? (void*) &fetch_val_l : (void*) &fetch_val_i /* out */,
prem /* dst */, type, rop, proc);
gmr_flush(dst_mreg, proc, 0); /* it's a round trip so w.r.t. flush, local=remote */
if (is_long)
*(long*) ploc = fetch_val_l;
else
*(int*) ploc = fetch_val_i;
}
return 0;
}
