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

Added OMP Barrier in ggml.c to avoid atomic operations

Browse source
This commit is contained in:
Abhishek Nair 2024-06-19 10:39:23 +05:30
parent 623494a478
commit 4147a04581
1 changed files with 34 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

37
ggml.c
View file

@ -19033,8 +19033,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
state->ec = GGML_STATUS_ABORTED; state->ec = GGML_STATUS_ABORTED;
return 0; return 0;
} }
#ifdef GGML_USE_OPENMP
if (state->ith == 0) {
#else
if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
#endif
// all other threads are finished and spinning // all other threads are finished and spinning
// do finalize and init here so we don't have synchronize again // do finalize and init here so we don't have synchronize again
struct ggml_compute_params params = { struct ggml_compute_params params = {
@ -19094,6 +19097,15 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
} }
task_phase = GGML_TASK_TYPE_INIT; task_phase = GGML_TASK_TYPE_INIT;
#ifdef GGML_USE_OPENMP
state->shared->n_active = n_threads;
state->shared->node_n = node_n;
state->shared->node_task = task_phase;
}
#pragma omp barrier
node_n = state->shared->node_n;
task_phase = state->shared->node_task;
#else
atomic_store(&state->shared->n_active, n_threads); atomic_store(&state->shared->n_active, n_threads);
atomic_store(&state->shared->node_n, node_n); atomic_store(&state->shared->node_n, node_n);
atomic_store(&state->shared->node_task, task_phase); atomic_store(&state->shared->node_task, task_phase);
@ -19101,6 +19113,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
ggml_graph_compute_thread_sync_node(&node_n, state, false); ggml_graph_compute_thread_sync_node(&node_n, state, false);
ggml_graph_compute_thread_sync_task(&task_phase, state, false); ggml_graph_compute_thread_sync_task(&task_phase, state, false);
} }
#endif
// check if we should stop // check if we should stop
if (node_n >= cgraph->n_nodes) break; if (node_n >= cgraph->n_nodes) break;
@ -19122,7 +19135,15 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
ggml_compute_forward(&params, node, state); ggml_compute_forward(&params, node, state);
} }
} }
#ifdef GGML_USE_OPENMP
if (state->ith == 0) {
task_phase = GGML_TASK_TYPE_COMPUTE;
state->shared->n_active = n_threads;
state->shared->node_task = task_phase;
}
#pragma omp barrier
task_phase = state->shared->node_task;
#else
if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
task_phase = GGML_TASK_TYPE_COMPUTE; task_phase = GGML_TASK_TYPE_COMPUTE;
atomic_store(&state->shared->n_active, n_threads); atomic_store(&state->shared->n_active, n_threads);
@ -19137,12 +19158,21 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT; 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); ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield);
} }
#endif
if (state->ith < n_tasks) { if (state->ith < n_tasks) {
params.type = GGML_TASK_TYPE_COMPUTE; params.type = GGML_TASK_TYPE_COMPUTE;
ggml_compute_forward(&params, node, state); ggml_compute_forward(&params, node, state);
} }
#ifdef GGML_USE_OPENMP
if (state->ith == 0) {
task_phase = GGML_TASK_TYPE_FINALIZE;
state->shared->n_active = n_threads;
state->shared->node_task = task_phase;
}
#pragma omp barrier
task_phase = state->shared->node_task;
#else
if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
task_phase = GGML_TASK_TYPE_FINALIZE; task_phase = GGML_TASK_TYPE_FINALIZE;
atomic_store(&state->shared->n_active, n_threads); atomic_store(&state->shared->n_active, n_threads);
@ -19151,6 +19181,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
else { else {
ggml_graph_compute_thread_sync_task(&task_phase, state, false); ggml_graph_compute_thread_sync_task(&task_phase, state, false);
} }
#endif
} }
return 0; return 0;