Anatolii Rozanov
d0c8380650
* Add functional test for barrier_all_on_stream
* Add rocshmem_barrier_all_on_stream support for GDA and RO backends
Implements rocshmem_barrier_all_on_stream operation for
GPU Direct Access and Reverse Offload backends.
Previously, rocshmem_barrier_all_on_stream was only supported for IPC backend.
* Add functional test for rocshmem_broadcastmem_on_stream
* Add host-side rocshmem_broadcastmem_on_stream API
Implement stream-based broadcast collective operation
- Add rocshmem_broadcastmem_on_stream host API and kernel implementation
- Add functional test TeamBroadcastmemOnStreamTester with multi-stream
support and correctness verification
- Use per-workgroup contexts to avoid contention across parallel streams
API:
rocshmem_broadcastmem_on_stream(team, dest, source, nelems, pe_root, stream)
* Add functional test for rocshmem_getmem_on_stream
* Add host-side rocshmem_getmem_on_stream API
Implement stream-based point-to-point RMA get operation
- Add rocshmem_getmem_on_stream host API and kernel implementation
- Support for asynchronous getmem operations on HIP streams
- Add backend support for GDA, RO, and IPC contexts
- Use work-group collective getmem for efficient memory transfer
API:
rocshmem_getmem_on_stream(dest, source, nelems, pe, stream)
(AI Assist)
* Add host-side rocshmem_putmem_on_stream API
- Add rocshmem_putmem_on_stream for asynchronous remote writes
- Support for concurrent RMA operations on HIP streams
- Add backend support for GDA, RO, and IPC contexts
- Use work-group device collective operation
API:
rocshmem_putmem_on_stream(dest, source, bytes, pe, stream)
(AI Assist)
* Add functional test for rocshmem_putmem_on_stream
* Add host-side rocshmem_putmem_signal_on_stream API
Enables asynchronous putmem operations with signaling on HIP streams.
The implementation includes:
- Kernel wrapper rocshmem_putmem_signal_kernel
- Host interface putmem_signal_on_stream method
- Context layer support across all backends (IPC, GDA, RO)
- Public API
Function signature:
void rocshmem_putmem_signal_on_stream(void *dest, const void *source,
size_t bytes, uint64_t *sig_addr,
uint64_t signal, int sig_op,
int pe, hipStream_t stream);
* Add functional test for rocshmem_putmem_signal_on_stream
* Add host-side rocshmem_signal_wait_until_on_stream API
Enables asynchronous signal wait operations on HIP streams.
The implementation includes:
- Kernel wrapper rocshmem_signal_wait_until_kernel
- Host interface signal_wait_until_on_stream method
- Context layer support across all backends (IPC, GDA, RO)
- Native uint64_t support in wait_until API (generated from P2P_SYNC.py)
Function signature:
void rocshmem_signal_wait_until_on_stream(uint64_t *sig_addr, int cmp,
uint64_t cmp_value,
hipStream_t stream);
(AI Assist)
* Add functional test for rocshmem_signal_wait_until_on_stream
* Add documentation for stream API functions
This commit adds API documentation for the following host-side
stream functions:
- rocshmem_barrier_all_on_stream (collective routines)
- rocshmem_broadcastmem_on_stream (collective routines)
- rocshmem_getmem_on_stream (RMA operations)
- rocshmem_putmem_on_stream (RMA operations)
- rocshmem_putmem_signal_on_stream (signaling operations)
- rocshmem_signal_wait_until_on_stream (point-to-point sync)
The documentation includes function signatures, parameter descriptions,
and detailed explanations of asynchronous behavior and stream handling.
(AI Assist)
* Rename "bytes" -> "nelems"
* Add "_TEST_" to the variables used in tests
* Remove incorrect hipStreamDefault usage
hipStreamDefault is not a default stream. This is a flag.
If stream == nullptr, then just pass it to kernel. It will launch the kernel on the default stream
815 lines
33 KiB
C++
815 lines
33 KiB
C++
/******************************************************************************
|
|
* Copyright (c) Advanced Micro Devices, Inc. All rights reserved.
|
|
*
|
|
* SPDX-License-Identifier: MIT
|
|
*
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
* of this software and associated documentation files (the "Software"), to
|
|
* deal in the Software without restriction, including without limitation the
|
|
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
|
|
* sell copies of the Software, and to permit persons to whom the Software is
|
|
* furnished to do so, subject to the following conditions:
|
|
*
|
|
* The above copyright notice and this permission notice shall be included in
|
|
* all copies or substantial portions of the Software.
|
|
*
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
|
* IN THE SOFTWARE.
|
|
*****************************************************************************/
|
|
|
|
#ifndef LIBRARY_INCLUDE_ROCSHMEM_COLL_HPP
|
|
#define LIBRARY_INCLUDE_ROCSHMEM_COLL_HPP
|
|
|
|
namespace rocshmem {
|
|
|
|
/**
|
|
* @name SHMEM_ALLTOALL
|
|
* @brief Exchanges a fixed amount of contiguous data blocks between all pairs
|
|
* of PEs participating in the collective routine.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] team The team participating in the collective.
|
|
* @param[in] dest Destination address. Must be an address on the
|
|
* symmetric heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric
|
|
heap.
|
|
* @param[in] nelems Number of data blocks transferred per pair of PEs.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_float_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest,
|
|
const float *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_double_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest,
|
|
const double *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_char_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, char *dest,
|
|
const char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_schar_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, signed char *dest,
|
|
const signed char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_short_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest,
|
|
const short *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_int_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest,
|
|
const int *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_long_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest,
|
|
const long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_longlong_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest,
|
|
const long long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uchar_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned char *dest,
|
|
const unsigned char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ushort_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned short *dest,
|
|
const unsigned short *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uint_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned int *dest,
|
|
const unsigned int *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulong_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long *dest,
|
|
const unsigned long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulonglong_alltoall_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long long *dest,
|
|
const unsigned long long *source, int nelems);
|
|
|
|
|
|
/**
|
|
* @name SHMEM_BROADCAST
|
|
* @brief Perform a broadcast between PEs in the active set. The caller
|
|
* is blocked until the broadcase completes.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] dest Destination address. Must be an address on the
|
|
* symmetric heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric
|
|
heap.
|
|
* @param[in] nelement Size of the buffer to participate in the broadcast.
|
|
* @param[in] PE_root Zero-based ordinal of the PE, with respect to the
|
|
active set, from which the data is copied
|
|
* @param[in] PE_start PE to start the reduction.
|
|
* @param[in] logPE_stride Stride of PEs participating in the reduction.
|
|
* @param[in] PE_size Number PEs participating in the reduction.
|
|
* @param[in] pSync Temporary sync buffer provided to ROCSHMEM. Must
|
|
be of size at least ROCSHMEM_REDUCE_SYNC_SIZE.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_float_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest,
|
|
const float *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_float_broadcast(
|
|
rocshmem_ctx_t ctx, float *dest, const float *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_float_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest,
|
|
const float *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_double_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest,
|
|
const double *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_double_broadcast(
|
|
rocshmem_ctx_t ctx, double *dest, const double *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_double_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest,
|
|
const double *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_char_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, char *dest,
|
|
const char *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_char_broadcast(
|
|
rocshmem_ctx_t ctx, char *dest, const char *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_char_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, char *dest,
|
|
const char *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_schar_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, signed char *dest,
|
|
const signed char *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_schar_broadcast(
|
|
rocshmem_ctx_t ctx, signed char *dest, const signed char *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_schar_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, signed char *dest,
|
|
const signed char *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_short_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest,
|
|
const short *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_short_broadcast(
|
|
rocshmem_ctx_t ctx, short *dest, const short *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_short_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest,
|
|
const short *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_int_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest,
|
|
const int *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_int_broadcast(
|
|
rocshmem_ctx_t ctx, int *dest, const int *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_int_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest,
|
|
const int *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_long_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest,
|
|
const long *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_long_broadcast(
|
|
rocshmem_ctx_t ctx, long *dest, const long *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_long_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest,
|
|
const long *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_longlong_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest,
|
|
const long long *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_longlong_broadcast(
|
|
rocshmem_ctx_t ctx, long long *dest, const long long *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_longlong_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest,
|
|
const long long *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uchar_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned char *dest,
|
|
const unsigned char *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_uchar_broadcast(
|
|
rocshmem_ctx_t ctx, unsigned char *dest, const unsigned char *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_uchar_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned char *dest,
|
|
const unsigned char *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ushort_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned short *dest,
|
|
const unsigned short *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_ushort_broadcast(
|
|
rocshmem_ctx_t ctx, unsigned short *dest, const unsigned short *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_ushort_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned short *dest,
|
|
const unsigned short *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uint_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned int *dest,
|
|
const unsigned int *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_uint_broadcast(
|
|
rocshmem_ctx_t ctx, unsigned int *dest, const unsigned int *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_uint_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned int *dest,
|
|
const unsigned int *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulong_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long *dest,
|
|
const unsigned long *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_ulong_broadcast(
|
|
rocshmem_ctx_t ctx, unsigned long *dest, const unsigned long *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_ulong_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long *dest,
|
|
const unsigned long *source, int nelems, int pe_root);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulonglong_broadcast_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long long *dest,
|
|
const unsigned long long *source, int nelems, int pe_root);
|
|
__host__ void rocshmem_ctx_ulonglong_broadcast(
|
|
rocshmem_ctx_t ctx, unsigned long long *dest, const unsigned long long *source,
|
|
int nelems, int pe_root, int pe_start, int log_pe_stride,
|
|
int pe_size, long *p_sync);
|
|
__host__ void rocshmem_ctx_ulonglong_broadcast(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long long *dest,
|
|
const unsigned long long *source, int nelems, int pe_root);
|
|
|
|
|
|
/**
|
|
* @name SHMEM_FCOLLECT
|
|
* @brief Concatenates blocks of data from multiple PEs to an array in every
|
|
* PE participating in the collective routine.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] team The team participating in the collective.
|
|
* @param[in] dest Destination address. Must be an address on the
|
|
* symmetric heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric
|
|
heap.
|
|
* @param[in] nelems Number of data blocks in source array.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_float_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest,
|
|
const float *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_double_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest,
|
|
const double *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_char_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, char *dest,
|
|
const char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_schar_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, signed char *dest,
|
|
const signed char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_short_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest,
|
|
const short *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_int_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest,
|
|
const int *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_long_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest,
|
|
const long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_longlong_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest,
|
|
const long long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uchar_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned char *dest,
|
|
const unsigned char *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ushort_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned short *dest,
|
|
const unsigned short *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_uint_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned int *dest,
|
|
const unsigned int *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulong_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long *dest,
|
|
const unsigned long *source, int nelems);
|
|
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_ulonglong_fcollect_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, unsigned long long *dest,
|
|
const unsigned long long *source, int nelems);
|
|
|
|
|
|
/**
|
|
* @name SHMEM_REDUCTIONS
|
|
* @brief Perform an allreduce between PEs in the active set. The caller
|
|
* is blocked until the reduction completes.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] team The team participating in the collective.
|
|
* @param[in] dest Destination address. Must be an address on the
|
|
* symmetric heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric
|
|
heap.
|
|
* @param[in] nreduce Size of the buffer to participate in the reduction.
|
|
*
|
|
* @return int (Zero on successful local completion. Nonzero otherwise.)
|
|
*/
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_or_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_or_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_and_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_and_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_short_xor_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_short_xor_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, short *dest, const short *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_or_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_or_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_and_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_and_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_int_xor_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_int_xor_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, int *dest, const int *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_or_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_or_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_and_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_and_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_long_xor_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_long_xor_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long *dest, const long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_or_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_or_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_and_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_and_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_longlong_xor_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_longlong_xor_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, long long *dest, const long long *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_float_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_float_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_float_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_float_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_float_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_float_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_float_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_float_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, float *dest, const float *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_double_sum_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_double_sum_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_double_min_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_double_min_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_double_max_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_double_max_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
|
|
__device__ ATTR_NO_INLINE int rocshmem_ctx_double_prod_reduce_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
__host__ int rocshmem_ctx_double_prod_reduce(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team, double *dest, const double *source,
|
|
int nreduce);
|
|
|
|
/**
|
|
* @brief kernel for performing a barrier synchronization.
|
|
* Caller enqueues the kernel on given stream
|
|
*
|
|
* @return void
|
|
*/
|
|
__global__ ATTR_NO_INLINE void rocshmem_barrier_all_kernel();
|
|
|
|
/**
|
|
* @brief kernel for performing an alltoall collective operation.
|
|
* Caller enqueues the kernel on given stream
|
|
*
|
|
* @param[in] team The team participating in the collective.
|
|
* @param[in] dest Destination address. Must be an address on the symmetric
|
|
* heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric heap.
|
|
* @param[in] size Number of bytes to transfer per pair of PEs.
|
|
*
|
|
* @return void
|
|
*/
|
|
__global__ ATTR_NO_INLINE void rocshmem_alltoallmem_kernel(rocshmem_team_t team,
|
|
void *dest,
|
|
const void *source,
|
|
size_t size);
|
|
|
|
/**
|
|
* @brief kernel for performing a broadcast collective operation.
|
|
* Caller enqueues the kernel on given stream
|
|
*
|
|
* @param[in] team The team participating in the collective.
|
|
* @param[in] dest Destination address. Must be an address on the symmetric
|
|
* heap.
|
|
* @param[in] source Source address. Must be an address on the symmetric heap.
|
|
* @param[in] nelems Number of bytes to broadcast.
|
|
* @param[in] pe_root Root PE (relative to team) from which to broadcast.
|
|
*
|
|
* @return void
|
|
*/
|
|
__global__ ATTR_NO_INLINE void rocshmem_broadcastmem_kernel(
|
|
rocshmem_team_t team, void *dest, const void *source, size_t nelems,
|
|
int pe_root);
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the system.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be invoked by a single thread within the PE.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_barrier_all();
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the system.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be called as a wave-front collective.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_barrier_all_wave();
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the system.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_barrier_all_wg();
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the team.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be invoked by a single thread within the PE.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
*
|
|
* @param[in] team The team on which to perform barrier synchronization
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ void rocshmem_ctx_barrier(rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the team.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be called as a wave-front collective.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
*
|
|
* @param[in] team The team on which to perform barrier synchronization
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ void rocshmem_ctx_barrier_wave(rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
/**
|
|
* @brief perform a collective barrier between all PEs in the team.
|
|
* The caller is blocked until the barrier is resolved.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
*
|
|
* @param[in] team The team on which to perform barrier synchronization
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ void rocshmem_ctx_barrier_wg(rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_sync_all only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be invoked by a single thread within the PE.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_sync_all();
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_sync_all only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be called as a wave-front collective.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_sync_all_wave();
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_sync_all only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_sync_all_wg();
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_team_sync only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be invoked by a single thread within the PE.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
* @param[in] team Handle of the team being synchronized
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_sync(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_team_sync only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be called as a wave-front collective.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
* @param[in] team Handle of the team being synchronized
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_sync_wave(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
/**
|
|
* @brief registers the arrival of a PE at a barrier.
|
|
* The caller is blocked until the synchronization is resolved.
|
|
*
|
|
* In contrast with the shmem_barrier_all routine, shmem_team_sync only ensures
|
|
* completion and visibility of previously issued memory stores and does not
|
|
* ensure completion of remote memory updates issued via OpenSHMEM routines.
|
|
*
|
|
* This function must be called as a work-group collective.
|
|
*
|
|
* @param[in] handle GPU side handle.
|
|
* @param[in] team Handle of the team being synchronized
|
|
*
|
|
* @return void
|
|
*/
|
|
__device__ ATTR_NO_INLINE void rocshmem_ctx_sync_wg(
|
|
rocshmem_ctx_t ctx, rocshmem_team_t team);
|
|
|
|
} // namespace rocshmem
|
|
|
|
#endif // LIBRARY_INCLUDE_ROCSHMEM_COLL_HPP
|