vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

threadpool: simplify threadpool init logic and fix main thread affinity application

Browse source 
Most of the init code is now exactly the same between threadpool and openmp.
This commit is contained in:
Max Krasnyansky 2024-08-24 17:35:34 -07:00 • committed by fmz
parent 8008463aee
commit 49ac51f2a3
1 changed files with 26 additions and 49 deletions
Download patch file Download diff file Expand all files Collapse all files

75
ggml/src/ggml.c
View file

@ -19191,7 +19191,6 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
struct ggml_compute_threadpool * threadpool = state->threadpool; struct ggml_compute_threadpool * threadpool = state->threadpool;
ggml_thread_apply_thread_priority(threadpool->prio); ggml_thread_apply_thread_priority(threadpool->prio);
if (ggml_thread_cpumask_is_valid(state->cpumask)) { if (ggml_thread_cpumask_is_valid(state->cpumask)) {
ggml_thread_apply_affinity(state->cpumask); ggml_thread_apply_affinity(state->cpumask);
} }
@ -19296,51 +19295,35 @@ static struct ggml_compute_threadpool * ggml_create_threadpool_impl(
threadpool->ec = GGML_STATUS_SUCCESS; threadpool->ec = GGML_STATUS_SUCCESS;
} }
#ifndef GGML_USE_OPENMP // Allocate and init workers state
ggml_mutex_init(&threadpool->mutex); const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads;
ggml_cond_init(&threadpool->cond); struct ggml_compute_state * workers = GGML_ALIGNED_MALLOC(workers_size);
#endif // GGML_USE_OPENMP
struct ggml_compute_state * workers = memset(workers, 0, workers_size);
GGML_ALIGNED_MALLOC(sizeof(struct ggml_compute_state) * tpp->n_threads); for (int j = 0; j < tpp->n_threads; j++) {
workers[j].threadpool = threadpool;
workers[j].ith = j;
}
threadpool->workers = workers; threadpool->workers = workers;
#ifdef GGML_USE_OPENMP #ifndef GGML_USE_OPENMP
for (int j = 0; j < tpp->n_threads; j++) { ggml_mutex_init(&threadpool->mutex);
workers[j] = (struct ggml_compute_state) { ggml_cond_init(&threadpool->cond);
.threadpool = threadpool,
.ith = j // Spin the threads for all workers, and update CPU placements.
}; // Place the main thread last (towards the higher numbered CPU cores).
}
#else // Not using OPENMP
int32_t cpumask_iter = 0; int32_t cpumask_iter = 0;
ggml_thread_apply_process_priority(tpp->prio); for (int j = 1; j < tpp->n_threads; j++) {
ggml_thread_apply_thread_priority(tpp->prio);
for (int j = 0; j < tpp->n_threads; j++) {
workers[j] = (struct ggml_compute_state) {
.thrd = 0,
.threadpool = threadpool,
.ith = j,
.last_graph = 0,
.pending = false
};
ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter); ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter);
// Spin threads for all secondary workers int32_t rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_secondary_thread, &workers[j]);
if (j > 0) { GGML_ASSERT(rc == 0);
int32_t rc = ggml_thread_create(
&workers[j].thrd,
NULL,
ggml_graph_compute_secondary_thread,
&workers[j]
);
GGML_ASSERT(rc == 0);
}
} }
ggml_thread_cpumask_next(tpp->cpumask, workers[0].cpumask, tpp->strict_cpu, &cpumask_iter);
#endif // GGML_USE_OPENMP #endif // GGML_USE_OPENMP
return threadpool; return threadpool;
@ -19391,22 +19374,16 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
threadpool->n_threads_cur = n_threads; threadpool->n_threads_cur = n_threads;
} }
struct ggml_compute_state worker = { ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
.ith = omp_get_thread_num(),
.threadpool = threadpool,
};
ggml_graph_compute_thread(&worker);
} }
} else { } else {
struct ggml_compute_state worker = { ggml_graph_compute_thread(&threadpool->workers[0]);
.ith = 0,
.threadpool = threadpool,
};
ggml_graph_compute_thread(&worker);
} }
#else #else
// Update main thread affinity to match the current threadpool // Update main thread prio and affinity to match the current threadpool
if (!ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) { ggml_thread_apply_process_priority(threadpool->prio);
ggml_thread_apply_thread_priority(threadpool->prio);
if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) {
ggml_thread_apply_affinity(threadpool->workers[0].cpumask); ggml_thread_apply_affinity(threadpool->workers[0].cpumask);
} }