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

Avoid INIT synchronization barrier when possible

Browse source 
This change makes inference go ~5% faster for me.
This commit is contained in:
Justine Tunney 2024-05-22 00:31:28 -07:00
parent 7deec14bd9
commit 8435ab0ae8
No known key found for this signature in database
GPG key ID: 52965314629936D4
1 changed files with 15 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

28
ggml.c
View file

@ -20013,19 +20013,21 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
} }
} }
if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { if (GGML_OP_HAS_INIT[node->op]) {
task_phase = GGML_TASK_TYPE_COMPUTE; if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
atomic_store(&state->shared->n_active, n_threads); task_phase = GGML_TASK_TYPE_COMPUTE;
atomic_store(&state->shared->node_task, task_phase); atomic_store(&state->shared->n_active, n_threads);
} atomic_store(&state->shared->node_task, task_phase);
else { }
// TODO: this sched_yield can have significant impact on the performance - either positive or negative else {
// depending on the workload and the operating system. // TODO: this sched_yield can have significant impact on the performance - either positive or negative
// since it is not clear what is the best approach, it should potentially become user-configurable // depending on the workload and the operating system.
// ref: https://github.com/ggerganov/ggml/issues/291 // since it is not clear what is the best approach, it should potentially become user-configurable
// UPD: adding the do_yield flag seems to resolve the issue universally // ref: https://github.com/ggerganov/ggml/issues/291
const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT; // UPD: adding the do_yield flag seems to resolve the issue universally
ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield); const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT;
ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield);
}
} }
if (state->ith < n_tasks) { if (state->ith < n_tasks) {