/* StarPU --- Runtime system for heterogeneous multicore architectures. * * Copyright (C) 2012,2013 Inria * Copyright (C) 2010-2013,2015-2017 CNRS * Copyright (C) 2010-2015 Université de Bordeaux * * 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. */ /* * This shows how to implement an spmd parallel StarPU task: scal_cpu_func is * called in parallel over several cores, and has to split the work accordingly. * This is a mere vector scaling example. */ /* gcc build: gcc -O2 -g vector_scal.c -o vector_scal $(pkg-config --cflags starpu-1.0) $(pkg-config --libs starpu-1.0) */ #include #include #include #define MIN(a,b) ((a)<(b)?(a):(b)) #define NX 204800 #define FPRINTF(ofile, fmt, ...) do { if (!getenv("STARPU_SSILENT")) {fprintf(ofile, fmt, ## __VA_ARGS__); }} while(0) #ifdef STARPU_QUICK_CHECK #define ITER 10 #else #define ITER 100 #endif static int get_first_element_rank(int nel, int rank, int nb_workers) { if(rank == 0) return 0; /* We get the number of bigger parts which stand before the part */ int nb_big_parts = MIN(nel % nb_workers, rank); return nb_big_parts * (nel / nb_workers + 1) + (rank - nb_big_parts) * (nel / nb_workers); } void scal_cpu_func(void *buffers[], void *_args) { int i; float *factor = _args, f = *factor; struct starpu_vector_interface *vector = buffers[0]; int n = STARPU_VECTOR_GET_NX(vector); float *val = (float *)STARPU_VECTOR_GET_PTR(vector); int nb_workers = starpu_combined_worker_get_size(); int rank = starpu_combined_worker_get_rank(); if (rank == 0) FPRINTF(stderr, "running task with %d CPUs.\n", starpu_combined_worker_get_size()); /* We add 1 to the (nel_total % nb_workers) first workers, thus we get an evenly splitted data. */ int nel_worker = (n / nb_workers) + ((rank < (n % nb_workers)) ? 1 : 0); int begin = get_first_element_rank(n, rank, nb_workers); for (i = 0; i < nel_worker; i++) { rank = i + begin; float v = val[rank]; int j; for (j = 0; j < 100; j++) v = v * f; val[rank] = v; } } static struct starpu_perfmodel vector_scal_model = { .type = STARPU_HISTORY_BASED, .symbol = "vector_scal_parallel" }; static struct starpu_codelet cl = { .modes = { STARPU_RW }, .type = STARPU_SPMD, .max_parallelism = INT_MAX, .cpu_funcs = {scal_cpu_func}, .cpu_funcs_name = {"scal_cpu_func"}, .nbuffers = 1, .model = &vector_scal_model, }; int main(void) { struct starpu_conf conf; float *vector; unsigned i; int ret; starpu_conf_init(&conf); conf.single_combined_worker = 1; conf.sched_policy_name = "pheft"; ret = starpu_init(&conf); if (ret == -ENODEV) return 77; STARPU_CHECK_RETURN_VALUE(ret, "starpu_init"); vector = malloc(NX*sizeof(*vector)); for (i = 0; i < NX; i++) vector[i] = (i+1.0f); FPRINTF(stderr, "BEFORE: First element was %f\n", vector[0]); FPRINTF(stderr, "BEFORE: Last element was %f\n", vector[NX-1]); starpu_data_handle_t vector_handle; starpu_vector_data_register(&vector_handle, STARPU_MAIN_RAM, (uintptr_t)vector, NX, sizeof(vector[0])); float factor = 1.001; for (i = 0; i < ITER; i++) { struct starpu_task *task = starpu_task_create(); task->cl = &cl; task->handles[0] = vector_handle; task->cl_arg = &factor; task->cl_arg_size = sizeof(factor); ret = starpu_task_submit(task); if (ret == -ENODEV) { ret = 77; break; } STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit"); } starpu_data_unregister(vector_handle); /* terminate StarPU, no task can be submitted after */ starpu_shutdown(); FPRINTF(stderr, "AFTER: First element is %f\n", vector[0]); FPRINTF(stderr, "AFTER: Last element is %f\n", vector[NX-1]); free(vector); return ret; }