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

ggml : minor style changes

Browse source
This commit is contained in:
Georgi Gerganov 2023-06-19 18:45:36 +03:00
parent 90a0e65c67
commit d0e3596350
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 16 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml.c
View file

@ -16342,6 +16342,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
/*.wsize =*/ cgraph->work ? ggml_nbytes(cgraph->work) : 0,
/*.wdata =*/ cgraph->work ? cgraph->work->data : NULL,
};
if (node_n != -1) {
/* FINALIZE */
struct ggml_tensor * node = state->shared->cgraph->nodes[node_n];
@ -16349,6 +16350,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
ggml_compute_forward(&params, node);
ggml_graph_compute_perf_stats_node(node, state->shared);
}
// distribute new work or execute it direct if 1T
while (++node_n < cgraph->n_nodes) {
GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
@ -16368,6 +16370,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
// they do something more efficient than spinning (?)
params.type = GGML_TASK_COMPUTE;
ggml_compute_forward(&params, node);
params.type = GGML_TASK_FINALIZE;
ggml_compute_forward(&params, node);
ggml_graph_compute_perf_stats_node(node, state->shared);
@ -16375,6 +16378,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
break;
}
}
atomic_store(&state->shared->n_active, n_threads);
atomic_store(&state->shared->node_n, node_n);
} else {
@ -16387,8 +16391,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
}
// check if we should stop
if (node_n >= cgraph->n_nodes) break;
struct ggml_tensor * node = cgraph->nodes[node_n];
/* COMPUTE */
struct ggml_tensor * node = cgraph->nodes[node_n];
struct ggml_compute_params params = {
/*.type =*/ GGML_TASK_COMPUTE,
/*.ith =*/ state->ith,
@ -16396,12 +16401,14 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
/*.wsize =*/ cgraph->work ? ggml_nbytes(cgraph->work) : 0,
/*.wdata =*/ cgraph->work ? cgraph->work->data : NULL,
};
if(state->ith < node->n_tasks) {
ggml_compute_forward(&params, node);
} else {
break;
}
}
return 0;
}
@ -16778,7 +16785,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
workers[0].ith = 0;
workers[0].shared = &state_shared;
const int64_t perf_start_cycles = ggml_perf_cycles();
const int64_t perf_start_time_us = ggml_perf_time_us();