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

24
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, /*.wsize =*/ cgraph->work ? ggml_nbytes(cgraph->work) : 0,
/*.wdata =*/ cgraph->work ? cgraph->work->data : NULL, /*.wdata =*/ cgraph->work ? cgraph->work->data : NULL,
}; };
if (node_n != -1) { if (node_n != -1) {
/* FINALIZE */ /* FINALIZE */
struct ggml_tensor * node = state->shared->cgraph->nodes[node_n]; struct ggml_tensor * node = state->shared->cgraph->nodes[node_n];
@ -16349,18 +16350,19 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
ggml_graph_compute_perf_stats_node(node, state->shared); ggml_graph_compute_perf_stats_node(node, state->shared);
} }
// distribute new work or execute it direct if 1T // distribute new work or execute it direct if 1T
while (++node_n < cgraph->n_nodes) { while (++node_n < cgraph->n_nodes) {
GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
struct ggml_tensor * node = cgraph->nodes[node_n]; struct ggml_tensor * node = cgraph->nodes[node_n];
state->shared->perf_node_start_cycles = ggml_perf_cycles(); state->shared->perf_node_start_cycles = ggml_perf_cycles();
state->shared->perf_node_start_time_us = ggml_perf_time_us(); state->shared->perf_node_start_time_us = ggml_perf_time_us();
/* INIT */ /* INIT */
params.type = GGML_TASK_INIT; params.type = GGML_TASK_INIT;
params.nth = node->n_tasks; params.nth = node->n_tasks;
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
if (node->n_tasks == 1) { if (node->n_tasks == 1) {
@ -16368,6 +16370,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
// they do something more efficient than spinning (?) // they do something more efficient than spinning (?)
params.type = GGML_TASK_COMPUTE; params.type = GGML_TASK_COMPUTE;
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
params.type = GGML_TASK_FINALIZE; params.type = GGML_TASK_FINALIZE;
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
ggml_graph_compute_perf_stats_node(node, state->shared); ggml_graph_compute_perf_stats_node(node, state->shared);
@ -16375,6 +16378,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
break; break;
} }
} }
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);
} else { } else {
@ -16387,8 +16391,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
} }
// check if we should stop // check if we should stop
if (node_n >= cgraph->n_nodes) break; if (node_n >= cgraph->n_nodes) break;
struct ggml_tensor * node = cgraph->nodes[node_n];
/* COMPUTE */ /* COMPUTE */
struct ggml_tensor * node = cgraph->nodes[node_n];
struct ggml_compute_params params = { struct ggml_compute_params params = {
/*.type =*/ GGML_TASK_COMPUTE, /*.type =*/ GGML_TASK_COMPUTE,
/*.ith =*/ state->ith, /*.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, /*.wsize =*/ cgraph->work ? ggml_nbytes(cgraph->work) : 0,
/*.wdata =*/ cgraph->work ? cgraph->work->data : NULL, /*.wdata =*/ cgraph->work ? cgraph->work->data : NULL,
}; };
if(state->ith < node->n_tasks) { if(state->ith < node->n_tasks) {
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
} else { } else {
break; break;
} }
} }
return 0; return 0;
} }
@ -16409,12 +16416,12 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
const int n_threads = cgraph->n_threads; const int n_threads = cgraph->n_threads;
struct ggml_compute_state_shared state_shared = { struct ggml_compute_state_shared state_shared = {
/*.cgraph =*/ cgraph, /*.cgraph =*/ cgraph,
/*.perf_node_start_cycles =*/ 0, /*.perf_node_start_cycles =*/ 0,
/*.perf_node_start_time_us =*/ 0, /*.perf_node_start_time_us =*/ 0,
/*.n_threads =*/ n_threads, /*.n_threads =*/ n_threads,
/*.n_active =*/ n_threads, /*.n_active =*/ n_threads,
/*.node_n =*/ -1, /*.node_n =*/ -1,
}; };
struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads); struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads);
@ -16778,7 +16785,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
workers[0].ith = 0; workers[0].ith = 0;
workers[0].shared = &state_shared; workers[0].shared = &state_shared;
const int64_t perf_start_cycles = ggml_perf_cycles(); const int64_t perf_start_cycles = ggml_perf_cycles();
const int64_t perf_start_time_us = ggml_perf_time_us(); const int64_t perf_start_time_us = ggml_perf_time_us();