/* StarPU --- Runtime system for heterogeneous multicore architectures. * * Copyright (C) 2009-2021 Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria * Copyright (C) 2013 Thibaut Lambert * Copyright (C) 2016 Uppsala University * * StarPU is free software; you can redistribute it and/or modify * it under the terms of the GNU Lesser General Public License as published by * the Free Software Foundation; either version 2.1 of the License, or (at * your option) any later version. * * StarPU is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. * * See the GNU Lesser General Public License in COPYING.LGPL for more details. */ #ifndef __STARPU_WORKER_H__ #define __STARPU_WORKER_H__ #include #include #include #include #ifdef STARPU_HAVE_HWLOC #include #endif #ifdef __cplusplus extern "C" { #endif /** @defgroup API_Workers_Properties Workers’ Properties @{ */ /** Memory node Type */ enum starpu_node_kind { STARPU_UNUSED=0, STARPU_CPU_RAM=1, STARPU_CUDA_RAM=2, STARPU_OPENCL_RAM=3, STARPU_FPGA_RAM=4, STARPU_DISK_RAM=5, STARPU_MPI_MS_RAM=6, STARPU_MAX_RAM=6 }; /** Worker Architecture Type The value 4 which was used by the driver SCC is no longer used as renumbering workers would make unusable old performance model files. */ enum starpu_worker_archtype { STARPU_CPU_WORKER=0, /**< CPU core */ STARPU_CUDA_WORKER=1, /**< NVIDIA CUDA device */ STARPU_OPENCL_WORKER=2, /**< OpenCL device */ STARPU_FPGA_WORKER=4, /**< FPGA device */ STARPU_MPI_MS_WORKER=5, /**< MPI Slave device */ STARPU_NARCH = 6, /**< Number of arch types */ STARPU_ANY_WORKER=255 /**< any worker, used in the hypervisor */ }; /** Structure needed to iterate on the collection */ struct starpu_sched_ctx_iterator { /** The index of the current worker in the collection, needed when iterating on the collection. */ int cursor; void *value; void *possible_value; char visited[STARPU_NMAXWORKERS]; int possibly_parallel; }; /** Types of structures the worker collection can implement */ enum starpu_worker_collection_type { STARPU_WORKER_TREE, /**< The collection is a tree */ STARPU_WORKER_LIST /**< The collection is an array */ }; /** A scheduling context manages a collection of workers that can be memorized using different data structures. Thus, a generic structure is available in order to simplify the choice of its type. Only the list data structure is available but further data structures(like tree) implementations are foreseen. */ struct starpu_worker_collection { /** The workerids managed by the collection */ int *workerids; void *collection_private; /** The number of workers in the collection */ unsigned nworkers; void *unblocked_workers; unsigned nunblocked_workers; void *masters; unsigned nmasters; char present[STARPU_NMAXWORKERS]; char is_unblocked[STARPU_NMAXWORKERS]; char is_master[STARPU_NMAXWORKERS]; /** The type of structure */ enum starpu_worker_collection_type type; /** Check if there is another element in collection */ unsigned (*has_next)(struct starpu_worker_collection *workers, struct starpu_sched_ctx_iterator *it); /** Return the next element in the collection */ int (*get_next)(struct starpu_worker_collection *workers, struct starpu_sched_ctx_iterator *it); /** Add a new element in the collection */ int (*add)(struct starpu_worker_collection *workers, int worker); /** Remove an element from the collection */ int (*remove)(struct starpu_worker_collection *workers, int worker); /** Initialize the collection */ void (*init)(struct starpu_worker_collection *workers); /** Deinitialize the colection */ void (*deinit)(struct starpu_worker_collection *workers); /** Initialize the cursor if there is one */ void (*init_iterator)(struct starpu_worker_collection *workers, struct starpu_sched_ctx_iterator *it); void (*init_iterator_for_parallel_tasks)(struct starpu_worker_collection *workers, struct starpu_sched_ctx_iterator *it, struct starpu_task *task); }; extern struct starpu_worker_collection starpu_worker_list; extern struct starpu_worker_collection starpu_worker_tree; /** Return the number of workers (i.e. processing units executing StarPU tasks). The return value should be at most \ref STARPU_NMAXWORKERS. */ unsigned starpu_worker_get_count(void); /** Return the number of CPUs controlled by StarPU. The return value should be at most \ref STARPU_MAXCPUS. */ unsigned starpu_cpu_worker_get_count(void); /** Return the number of CUDA devices controlled by StarPU. The return value should be at most \ref STARPU_MAXCUDADEVS. */ unsigned starpu_cuda_worker_get_count(void); /** Return the number of OpenCL devices controlled by StarPU. The return value should be at most \ref STARPU_MAXOPENCLDEVS. */ unsigned starpu_opencl_worker_get_count(void); /** Return the number of MPI Master Slave workers controlled by StarPU. */ unsigned starpu_mpi_ms_worker_get_count(void); /** Return the identifier of the current worker, i.e the one associated to the calling thread. The return value is either \c -1 if the current context is not a StarPU worker (i.e. when called from the application outside a task or a callback), or an integer between \c 0 and starpu_worker_get_count() - \c 1. */ int starpu_worker_get_id(void); unsigned _starpu_worker_get_id_check(const char *f, int l); /** Similar to starpu_worker_get_id(), but abort when called from outside a worker (i.e. when starpu_worker_get_id() would return \c -1). */ unsigned starpu_worker_get_id_check(void); #define starpu_worker_get_id_check() _starpu_worker_get_id_check(__FILE__, __LINE__) int starpu_worker_get_bindid(int workerid); void starpu_sched_find_all_worker_combinations(void); /** Return the type of processing unit associated to the worker \p id. The worker identifier is a value returned by the function starpu_worker_get_id()). The return value indicates the architecture of the worker: ::STARPU_CPU_WORKER for a CPU core, ::STARPU_CUDA_WORKER for a CUDA device, and ::STARPU_OPENCL_WORKER for a OpenCL device. The return value for an invalid identifier is unspecified. */ enum starpu_worker_archtype starpu_worker_get_type(int id); /** Return the number of workers of \p type. A positive (or NULL) value is returned in case of success, -EINVAL indicates that \p type is not valid otherwise. */ int starpu_worker_get_count_by_type(enum starpu_worker_archtype type); /** Get the list of identifiers of workers of \p type. Fill the array \p workerids with the identifiers of the \p workers. The argument \p maxsize indicates the size of the array \p workerids. The return value gives the number of identifiers that were put in the array. -ERANGE is returned is \p maxsize is lower than the number of workers with the appropriate type: in that case, the array is filled with the \p maxsize first elements. To avoid such overflows, the value of maxsize can be chosen by the means of the function starpu_worker_get_count_by_type(), or by passing a value greater or equal to \ref STARPU_NMAXWORKERS. */ unsigned starpu_worker_get_ids_by_type(enum starpu_worker_archtype type, int *workerids, unsigned maxsize); /** Return the identifier of the \p num -th worker that has the specified \p type. If there is no such worker, -1 is returned. */ int starpu_worker_get_by_type(enum starpu_worker_archtype type, int num); /** Return the identifier of the worker that has the specified \p type and device id \p devid (which may not be the n-th, if some devices are skipped for instance). If there is no such worker, \c -1 is returned. */ int starpu_worker_get_by_devid(enum starpu_worker_archtype type, int devid); /** Get the name of the worker \p id. StarPU associates a unique human readable string to each processing unit. This function copies at most the \p maxlen first bytes of the unique string associated to the worker \p id into the \p dst buffer. The caller is responsible for ensuring that \p dst is a valid pointer to a buffer of \p maxlen bytes at least. Calling this function on an invalid identifier results in an unspecified behaviour. */ void starpu_worker_get_name(int id, char *dst, size_t maxlen); /** Display on \p output the list (if any) of all the workers of the given \p type. */ void starpu_worker_display_names(FILE *output, enum starpu_worker_archtype type); /** Return the device id of the worker \p id. The worker should be identified with the value returned by the starpu_worker_get_id() function. In the case of a CUDA worker, this device identifier is the logical device identifier exposed by CUDA (used by the function \c cudaGetDevice() for instance). The device identifier of a CPU worker is the logical identifier of the core on which the worker was bound; this identifier is either provided by the OS or by the library hwloc in case it is available. */ int starpu_worker_get_devid(int id); int starpu_worker_get_mp_nodeid(int id); struct starpu_tree* starpu_workers_get_tree(void); unsigned starpu_worker_get_sched_ctx_list(int worker, unsigned **sched_ctx); /** Return whether worker \p workerid is currently blocked in a parallel task. */ unsigned starpu_worker_is_blocked_in_parallel(int workerid); unsigned starpu_worker_is_slave_somewhere(int workerid); /** Return worker \p type as a string. */ const char *starpu_worker_get_type_as_string(enum starpu_worker_archtype type); /** Return worker \p type as a string suitable for environment variable names (CPU, CUDA, etc.) */ const char *starpu_worker_get_type_as_env_var(enum starpu_worker_archtype type); int starpu_bindid_get_workerids(int bindid, int **workerids); int starpu_worker_get_devids(enum starpu_worker_archtype type, int *devids, int num); int starpu_worker_get_stream_workerids(unsigned devid, int *workerids, enum starpu_worker_archtype type); unsigned starpu_worker_get_sched_ctx_id_stream(unsigned stream_workerid); #ifdef STARPU_HAVE_HWLOC /** If StarPU was compiled with \c hwloc support, return a duplicate of the \c hwloc cpuset associated with the worker \p workerid. The returned cpuset is obtained from a \c hwloc_bitmap_dup() function call. It must be freed by the caller using \c hwloc_bitmap_free(). */ hwloc_cpuset_t starpu_worker_get_hwloc_cpuset(int workerid); /** If StarPU was compiled with \c hwloc support, return the \c hwloc object corresponding to the worker \p workerid. */ hwloc_obj_t starpu_worker_get_hwloc_obj(int workerid); #endif int starpu_memory_node_get_devid(unsigned node); /** Return the memory node associated to the current worker */ unsigned starpu_worker_get_local_memory_node(void); /** Return the identifier of the memory node associated to the worker identified by \p workerid. */ unsigned starpu_worker_get_memory_node(unsigned workerid); /** Return the number of memory nodes */ unsigned starpu_memory_nodes_get_count(void); /** Return in \p name the name of a memory node (NUMA 0, CUDA 0, etc.) \p size is the size of the \p name array. */ int starpu_memory_node_get_name(unsigned node, char *name, size_t size); /** Return the number of NUMA nodes used by StarPU */ int starpu_memory_nodes_get_numa_count(void); /** Return the identifier of the memory node associated to the NUMA node identified by \p osid by the Operating System. */ int starpu_memory_nodes_numa_id_to_devid(int osid); /** Return the Operating System identifier of the memory node whose StarPU identifier is \p id. */ int starpu_memory_nodes_numa_devid_to_id(unsigned id); /** Return the type of \p node as defined by ::starpu_node_kind. For example, when defining a new data interface, this function should be used in the allocation function to determine on which device the memory needs to be allocated. */ enum starpu_node_kind starpu_node_get_kind(unsigned node); /** Return the type of worker which operates on memory node kind \p node_kind */ enum starpu_worker_archtype starpu_memory_node_get_worker_archtype(enum starpu_node_kind node_kind); /** Return the type of memory node that arch type \p type operates on */ enum starpu_node_kind starpu_worker_get_memory_node_kind(enum starpu_worker_archtype type); /** @name Scheduling operations @{ */ /** Return \c !0 if current worker has a scheduling operation in progress, and \c 0 otherwise. */ int starpu_worker_sched_op_pending(void); /** Allow other threads and workers to temporarily observe the current worker state, even though it is performing a scheduling operation. Must be called by a worker before performing a potentially blocking call such as acquiring a mutex other than its own sched_mutex. This function increases \c state_relax_refcnt from the current worker. No more than UINT_MAX-1 nested starpu_worker_relax_on() calls should performed on the same worker. This function is automatically called by starpu_worker_lock() to relax the caller worker state while attempting to lock the target worker. */ void starpu_worker_relax_on(void); /** Must be called after a potentially blocking call is complete, to restore the relax state in place before the corresponding starpu_worker_relax_on(). Decreases \c state_relax_refcnt. Calls to starpu_worker_relax_on() and starpu_worker_relax_off() must be properly paired. This function is automatically called by starpu_worker_unlock() after the target worker has been unlocked. */ void starpu_worker_relax_off(void); /** Return \c !0 if the current worker \c state_relax_refcnt!=0 and \c 0 otherwise. */ int starpu_worker_get_relax_state(void); /** Acquire the sched mutex of \p workerid. If the caller is a worker, distinct from \p workerid, the caller worker automatically enters a relax state while acquiring the target worker lock. */ void starpu_worker_lock(int workerid); /** Attempt to acquire the sched mutex of \p workerid. Returns \c 0 if successful, \c !0 if \p workerid sched mutex is held or the corresponding worker is not in a relax state. If the caller is a worker, distinct from \p workerid, the caller worker automatically enters relax state if successfully acquiring the target worker lock. */ int starpu_worker_trylock(int workerid); /** Release the previously acquired sched mutex of \p workerid. Restore the relax state of the caller worker if needed. */ void starpu_worker_unlock(int workerid); /** Acquire the current worker sched mutex. */ void starpu_worker_lock_self(void); /** Release the current worker sched mutex. */ void starpu_worker_unlock_self(void); #ifdef STARPU_WORKER_CALLBACKS /** If StarPU was compiled with blocking drivers support and worker callbacks support enabled, allow to specify an external resource manager callback to be notified about workers going to sleep. */ void starpu_worker_set_going_to_sleep_callback(void (*callback)(unsigned workerid)); /** If StarPU was compiled with blocking drivers support and worker callbacks support enabled, allow to specify an external resource manager callback to be notified about workers waking-up. */ void starpu_worker_set_waking_up_callback(void (*callback)(unsigned workerid)); #endif /** @} */ /** @} */ /** @defgroup API_Parallel_Tasks Parallel Tasks @{ */ /** Return the number of different combined workers. */ unsigned starpu_combined_worker_get_count(void); unsigned starpu_worker_is_combined_worker(int id); /** Return the identifier of the current combined worker. */ int starpu_combined_worker_get_id(void); /** Return the size of the current combined worker, i.e. the total number of CPUS running the same task in the case of ::STARPU_SPMD parallel tasks, or the total number of threads that the task is allowed to start in the case of ::STARPU_FORKJOIN parallel tasks. */ int starpu_combined_worker_get_size(void); /** Return the rank of the current thread within the combined worker. Can only be used in ::STARPU_SPMD parallel tasks, to know which part of the task to work on. */ int starpu_combined_worker_get_rank(void); /** Register a new combined worker and get its identifier */ int starpu_combined_worker_assign_workerid(int nworkers, int workerid_array[]); /** Get the description of a combined worker */ int starpu_combined_worker_get_description(int workerid, int *worker_size, int **combined_workerid); /** Variant of starpu_worker_can_execute_task() compatible with combined workers */ int starpu_combined_worker_can_execute_task(unsigned workerid, struct starpu_task *task, unsigned nimpl); /** Initialise the barrier for the parallel task, and dispatch the task between the different workers of the given combined worker. */ void starpu_parallel_task_barrier_init(struct starpu_task *task, int workerid); /** Initialise the barrier for the parallel task, to be pushed to \p worker_size workers (without having to explicit a given combined worker). */ void starpu_parallel_task_barrier_init_n(struct starpu_task *task, int worker_size); /** @} */ #ifdef __cplusplus } #endif #endif /* __STARPU_WORKER_H__ */