SHMEM DEF

Enumerations

group Enumerations

Defines

ACLSHMEM_P2P_SYNC_TYPE_FUNC(FUNC)

Standard test Types and Names.

NAME	TYPE
float	float
int8	int8
int16	int16
int32	int32
int64	int64
uint8	uint8
uint16	uint16
uint32	uint32
uint64	uint64
char	char

Typedefs

using Result = int32_t

Enums

enum aclshmem_block_mode_t

Enumeration indicating whether the method is blocking or non-blocking.

Values:

enumerator NO_NBI: Non-blocking mode disabled (method is blocking)

enumerator NBI: Non-blocking mode enabled (method is non-blocking, NBI = Non-Blocking Interface)

enum aclshmem_error_code_t

Error code for the ACLSHMEM library.

Values:

enumerator ACLSHMEM_SUCCESS: Task execution was successful.

enumerator ACLSHMEM_INVALID_PARAM: There is a problem with the parameters.

enumerator ACLSHMEM_INVALID_VALUE: There is a problem with the range of the value of the parameter.

enumerator ACLSHMEM_SMEM_ERROR: There is a problem with SMEM.

enumerator ACLSHMEM_INNER_ERROR: This is a problem caused by an internal error.

enumerator ACLSHMEM_NOT_INITED: This is a problem caused by an uninitialization.

enumerator ACLSHMEM_BOOTSTRAP_ERROR: This is a problem with BOOTSTRAP.

enumerator ACLSHMEM_TIMEOUT_ERROR: This is a problem caused by TIMEOUT.

enumerator ACLSHMEM_MALLOC_FAILED: This is a problem when malloc.

enumerator ACLSHMEM_DL_FUNC_FAILED: This is a problem when dll func failed.

enumerator ACLSHMEM_INNER_TIMEOUT: This is a problem when inner timeout.

enumerator ACLSHMEM_UNDER_API_UNLOAD: This is a problem when under api lib load failed.

enumerator ACLSHMEM_NOT_SUPPORTED: This is a problem when function not supported.

enum aclshmemx_bootstrap_t

init flags

Values:

enumerator ACLSHMEMX_INIT_WITH_DEFAULT

enumerator ACLSHMEMX_INIT_WITH_MPI

enumerator ACLSHMEMX_INIT_WITH_UNIQUEID

enumerator ACLSHMEMX_INIT_MAX

enum aclshmemx_init_status_t

The state of the ACLSHMEM library initialization.

Values:

enumerator ACLSHMEM_STATUS_NOT_INITIALIZED: Uninitialized.

enumerator ACLSHMEM_STATUS_SHM_CREATED: Shared memory heap creation is complete.

enumerator ACLSHMEM_STATUS_IS_INITIALIZED: Initialization is complete.

enumerator ACLSHMEM_STATUS_INVALID: Invalid status code.

enum aclshmem_transport_t

Different transports supported by ACLSHMEM library.

Values:

enumerator ACLSHMEM_TRANSPORT_MTE: MTE Transport.

enumerator ACLSHMEM_TRANSPORT_ROCE: RDMA Transport (RoCE).

enumerator ACLSHMEM_TRANSPORT_SDMA: SDMA Transport.

enum aclshmemi_op_t

The state of the ACLSHMEM host OP type.

Values:

enumerator ACLSHMEMI_OP_PUT

enumerator ACLSHMEMI_OP_P

enumerator ACLSHMEMI_OP_PUT_SIGNAL

enumerator ACLSHMEMI_OP_GET

enumerator ACLSHMEMI_OP_G

enum aclshmem_team_index_t

Team’s index.

Values:

enumerator ACLSHMEM_TEAM_INVALID

enumerator ACLSHMEM_TEAM_WORLD

enum data_op_engine_type_t

Data operation engine type.

Values:

enumerator ACLSHMEM_DATA_OP_MTE

enumerator ACLSHMEM_DATA_OP_SDMA

enumerator ACLSHMEM_DATA_OP_ROCE

enum aclshmem_mem_type_t

Memory type of NPU or host.

Values:

enumerator HOST_SIDE

enumerator DEVICE_SIDE

enum aclshmem_signal_op_type_t

Signal operations, used by the signaler in peer-to-peer synchronization.

Values:

enumerator ACLSHMEM_SIGNAL_SET

enumerator ACLSHMEM_SIGNAL_ADD

enum aclshmem_cmp_op_type_t

Signal comparison operations, used by the signalee in peer-to-peer synchronization.

Values:

enumerator ACLSHMEM_CMP_EQ

enumerator ACLSHMEM_CMP_NE

enumerator ACLSHMEM_CMP_GT

enumerator ACLSHMEM_CMP_GE

enumerator ACLSHMEM_CMP_LT

enumerator ACLSHMEM_CMP_LE

Typedefs

group Typedef

Typedefs

typedef int aclshmem_team_t: A typedef of int.

typedef uint64_t aclshmemx_team_uniqueid_t: A typedef of uint64_t.

typedef int32_t aclshmemi_sync_bit[ACLSHMEMI_SYNCBIT_SIZE / sizeof(int32_t)]: ACLSHMEM synchronization bit array type, cache line-aligned storage structure for synchronization flags.

Macros

group Macros

Defines

ACLSHMEM_HOST_API: A macro that identifies a function on the host side.

ACLSHMEM_MAJOR_VERSION: Macros that define major version info of ACLSHMEM.

ACLSHMEM_MINOR_VERSION: Macros that define minor version info of ACLSHMEM.

ACLSHMEM_MAX_NAME_LEN: Maximum length of the name string in ACLSHMEM (including null terminator)

ACLSHMEM_VENDOR_MAJOR_VER: Macros that define vendor major version info of ACLSHMEM.

ACLSHMEM_VENDOR_MINOR_VER: Macros that define vendor minor version info of ACLSHMEM.

ACLSHMEM_VENDOR_PATCH_VER: Macros that define vendor patch version info of ACLSHMEM.

ACLSHMEM_MAX_IP_PORT_LEN: Maximum length of the IP and port string in ACLSHMEM (including null terminator)

ACLSHMEM_UNIQUEID_INITIALIZER: Initializer macro for the ACLSHMEM unique ID structure.

ACLSHMEM_DEVICE: A macro that identifies a function on the device side.

ACLSHMEM_MAX_PES: Maximum number of Processing Elements (PE) supported by ACLSHMEM (16384)

ACLSHMEM_MAX_TEAMS: Maximum number of teams supported by ACLSHMEM (2048)

ACLSHMEM_MAX_LOCAL_SIZE: Maximum capacity of ACLSHMEM local memory (40GB)

TEAM_CONFIG_PADDING: Padding bytes for the team configuration structure, used for memory alignment (48 bytes)

SCALAR_DATA_CACHELINE_SIZE: Size of the scalar data cache line (64 bytes), used for cache alignment optimization.

L2_CACHELINE_SIZE: Size of the L2 cache line (512 bytes), used for advanced cache alignment optimization.

ACLSHMEM_PAGE_SIZE: ACLSHMEM memory page size (2MB), complying with system memory paging specifications.

ACLSHMEM_SDMA_MAX_CHAN: Max number of SDMA channels.

ACLSHMEM_SDMA_FLAG_LENGTH: SDMA flag data length.

ALIGH_TO(size, page)

Memory address/size alignment macro.

Aligns the size up to an integer multiple of the page to avoid memory fragmentation and out-of-bounds access

Parameters:

size – Original size/address to be aligned
page – Target alignment granularity (usually page size)

ACLSHMEMI_SYNCBIT_SIZE: Memory size of the ACLSHMEM synchronization bit, consistent with the scalar data cache line size.

SYNC_LOG_MAX_PES: Number of log bits required for NPU-level synchronization (5 bits), calculated as: ceil(log₈(16384)) = 5.

ACLSHMEM_BARRIER_TG_DISSEM_KVAL: Task group dissemination coefficient (8) for ACLSHMEM barrier synchronization, used for synchronization algorithm optimization.

SYNC_ARRAY_SIZE: Total size of the NPU-level synchronization array, calculated from sync bit size, log bits and dissemination coefficient.

SYNC_COUNTER_SIZE: Memory size of the NPU-level synchronization counter, consistent with the sync bit size.

SYNC_POOL_SIZE: Total size of the NPU-level synchronization pool, expanding the sync array by the maximum number of teams.

SYNC_COUNTERS_SIZE: Total size of the NPU-level synchronization counter pool, expanding the counter by the maximum number of teams.

ACLSHMEM_MAX_AIV_PER_NPU: Maximum number of AIV (AI Core) supported per NPU (48)

ACLSHMEM_LOG_MAX_AIV_PER_NPU: Number of log bits required for core-level synchronization (6 bits), calculated as: ceil(log₂(48)) = 6.

ACLSHMEM_CORE_SYNC_POOL_SIZE: Total size of the core-level synchronization pool, calculated from AIV count, log bits and sync bit size.

ACLSHMEM_CORE_SYNC_COUNTER_SIZE: Memory size of the core-level synchronization counter, consistent with the sync bit size.

ACLSHMEM_EXTRA_SIZE_UNALIGHED: Unaligned size of the ACLSHMEM extra memory (only includes NPU sync pool)

ACLSHMEM_EXTRA_SIZE: Aligned size of the ACLSHMEM extra memory, aligned to the memory page size.

Structs

group Structs

Defines

ACLSHMEM_CYCLE_PROF_MAX_BLOCK: cycle profling max block nums

ACLSHMEM_CYCLE_PROF_FRAME_CNT: cycle profling frame count

Typedefs

typedef struct aclshmemx_init_attr_t aclshmemx_init_attr_t

typedef aclshmem_ub_config_t aclshmem_mte_config_t

Deprecated:: Use aclshmem_ub_config_t instead

typedef aclshmem_ub_config_t aclshmem_sdma_config_t

Deprecated:: Use aclshmem_ub_config_t instead

typedef aclshmem_ub_config_t aclshmem_rdma_config_t

Deprecated:: Use aclshmem_ub_config_t instead

struct aclshmem_init_optional_attr_t

#include <shmem_host_def.h>

Optional parameter for the attributes used for initialization.

int version: version
data_op_engine_type_t data_op_engine_type: data_op_engine_type
uint32_t shm_init_timeout: shm_init_timeout
uint32_t shm_create_timeout: shm_create_timeout
uint32_t control_operation_timeout: control_operation_timeout
int32_t sockFd: sock_fd for apply port in advance

struct aclshmemx_init_attr_t

#include <shmem_host_def.h>

Mandatory parameter for attributes used for initialization.

int my_pe: The pe of the current process.
int n_pes: The total pe number of all processes.
char ip_port[ACLSHMEM_MAX_IP_PORT_LEN]: The ip and port of the communication server. The port must not conflict with other modules and processes.
uint64_t local_mem_size: The size of shared memory currently occupied by current pe.
aclshmem_init_optional_attr_t option_attr: Optional Parameters.
void *comm_args: Parameters required for communication during the bootstrap phase when initializing different flags.

struct aclshmemx_uniqueid_t

#include <shmem_host_def.h>

Structure required for SHMEM unique ID (uid) initialization.

int32_t version: version.
int my_pe: The pe of the current process.
int n_pes: The total pe number of all processes.
char internal[ACLSHMEM_UNIQUE_ID_INNER_LEN]: Internal information of uid.

struct non_contiguous_copy_param

#include <shmem_def.h>

Non-Contiguous Datacopy Param.

uint32_t repeat: Data move times
uint32_t length: Data move unit length
uint32_t src_ld: Src data leading dimension. Interval between the head of the repeat and the head of the following repeat.
uint32_t dst_ld: Dst data leading dimension.

struct aclshmem_handle_t