root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

ggml : do not sched_yield when calling BLAS (#4761)

Browse Source 
* ggml : do not sched_yield when calling BLAS

ggml-ci

* ggml : fix do_yield logic

ggml-ci

* ggml : simplify do_yield logic

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-01-05 15:18:21 +02:00 committed by GitHub
parent 3681f22443
commit c1d7cb28d3
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 14 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
ggml.c
View File

@ -9704,10 +9704,10 @@ static void ggml_compute_forward_group_norm(
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
// helper function to determine if it is better to use BLAS or not
// for large matrices, BLAS is faster
static bool ggml_compute_forward_mul_mat_use_blas(
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
struct ggml_tensor * dst) {
static bool ggml_compute_forward_mul_mat_use_blas(struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
//const int64_t ne00 = src0->ne[0];
//const int64_t ne01 = src0->ne[1];
@ -9787,7 +9787,7 @@ static void ggml_compute_forward_mul_mat(
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
if (ggml_compute_forward_mul_mat_use_blas(dst)) {
if (params->ith != 0) {
return;
}
@ -16301,24 +16301,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
//n_tasks = MIN(n_threads, MAX(1, nr0/128));
//printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks);
#if defined(GGML_USE_CUBLAS)
if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) {
n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
}
#elif defined(GGML_USE_CLBLAST)
if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) {
n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
}
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
}
#endif
} break;
case GGML_OP_MUL_MAT_ID:
{
@ -16491,6 +16473,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
state->shared->node_n += 1;
return (thread_ret_t) GGML_EXIT_ABORTED;
}
if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
// all other threads are finished and spinning
// do finalize and init here so we don't have synchronize again
@ -16556,14 +16539,18 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
} else {
// wait for other threads to finish
const int last = node_n;
const bool do_yield = last < 0 || cgraph->nodes[last]->op == GGML_OP_MUL_MAT;
while (true) {
// TODO: this sched_yield can have significant impact on the performance - either positive or negative
// depending on the workload and the operating system.
// since it is not clear what is the best approach, it should potentially become user-configurable
// ref: https://github.com/ggerganov/ggml/issues/291
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
sched_yield();
#endif
// UPD: adding the do_yield flag seems to resolve the issue universally
if (do_yield) {
sched_yield();
}
node_n = atomic_load(&state->shared->node_n);
if (node_n != last) break;
@ -16642,7 +16629,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
} else
#endif
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
if (ggml_compute_forward_mul_mat_use_blas(node)) {
if (node->src[0]->type != GGML_TYPE_F32) {
// here we need memory just for single 2D matrix from src0
cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]);