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

avoid sending finalize op to thread pool if it does nothing

Browse source
This commit is contained in:
zrm 2023-05-21 18:11:03 -04:00
parent 0d23f8ce8d
commit 9d058c2096
1 changed files with 8 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml.c
View file

@ -3698,6 +3698,12 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"f(x,y)", "f(x,y)",
}; };
// only send finalize op to thread pool if it actually does something
// currently none of them?
static const bool GGML_OP_HAS_FINALIZE[GGML_OP_COUNT] = {
0
};
static_assert(GGML_OP_COUNT == 51, "GGML_OP_COUNT != 51"); static_assert(GGML_OP_COUNT == 51, "GGML_OP_COUNT != 51");
static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@ -14541,7 +14547,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} }
// FINALIZE // FINALIZE
if (node->n_tasks > 1) { if (node->n_tasks > 1 && GGML_OP_HAS_FINALIZE[node->op]) {
if (atomic_fetch_add(&state_shared.n_ready, 1) == n_threads - 1) { if (atomic_fetch_add(&state_shared.n_ready, 1) == n_threads - 1) {
atomic_store(&state_shared.has_work, false); atomic_store(&state_shared.has_work, false);
} }
@ -14577,7 +14583,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
// wait for thread pool // wait for thread pool
if (node->n_tasks > 1) { if (node->n_tasks > 1 && GGML_OP_HAS_FINALIZE[node->op]) {
if (atomic_fetch_add(&state_shared.n_ready, 1) == n_threads - 1) { if (atomic_fetch_add(&state_shared.n_ready, 1) == n_threads - 1) {
atomic_store(&state_shared.has_work, false); atomic_store(&state_shared.has_work, false);
} }