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

remove ifdef GGML_PERF; update fmt

Browse source
This commit is contained in:
mqy 2023-07-15 17:36:39 +08:00
parent 218ab9ef89
commit 5d03303bdc
1 changed files with 4 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml.c
View file

@ -4410,7 +4410,7 @@ void ggml_free(struct ggml_context * ctx) {
if (&g_state.contexts[i].context == ctx) { if (&g_state.contexts[i].context == ctx) {
g_state.contexts[i].used = false; g_state.contexts[i].used = false;
GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n", GGML_PRINT_DEBUG("%s: context %d has been freed. memory used = %zu\n",
__func__, i, ggml_used_mem(ctx)); __func__, i, ggml_used_mem(ctx));
if (ctx->mem_buffer_owned) { if (ctx->mem_buffer_owned) {
@ -16249,7 +16249,6 @@ struct ggml_compute_state {
struct ggml_compute_state_shared * shared; struct ggml_compute_state_shared * shared;
}; };
#ifdef GGML_PERF
static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const struct ggml_compute_state_shared * st) { static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const struct ggml_compute_state_shared * st) {
int64_t cycles_cur = ggml_perf_cycles() - st->perf_node_start_cycles; int64_t cycles_cur = ggml_perf_cycles() - st->perf_node_start_cycles;
int64_t time_us_cur = ggml_perf_time_us() - st->perf_node_start_time_us; int64_t time_us_cur = ggml_perf_time_us() - st->perf_node_start_time_us;
@ -16258,7 +16257,6 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const
node->perf_cycles += cycles_cur; node->perf_cycles += cycles_cur;
node->perf_time_us += time_us_cur; node->perf_time_us += time_us_cur;
} }
#endif
static thread_ret_t ggml_graph_compute_thread(void * data) { static thread_ret_t ggml_graph_compute_thread(void * data) {
struct ggml_compute_state * state = (struct ggml_compute_state *) data; struct ggml_compute_state * state = (struct ggml_compute_state *) data;
@ -16296,9 +16294,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
params.nth = n_tasks_arr[node_n]; params.nth = n_tasks_arr[node_n];
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
} }
#ifdef GGML_PERF
ggml_graph_compute_perf_stats_node(node, state->shared); ggml_graph_compute_perf_stats_node(node, state->shared);
#endif
} }
// distribute new work or execute it direct if 1T // distribute new work or execute it direct if 1T
@ -16307,10 +16303,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
struct ggml_tensor * node = cgraph->nodes[node_n]; struct ggml_tensor * node = cgraph->nodes[node_n];
const int n_tasks = n_tasks_arr[node_n]; const int n_tasks = n_tasks_arr[node_n];
#ifdef GGML_PERF
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();
#endif
params.nth = n_tasks; params.nth = n_tasks;
/* INIT */ /* INIT */
@ -16329,9 +16325,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
params.type = GGML_TASK_FINALIZE; params.type = GGML_TASK_FINALIZE;
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
} }
#ifdef GGML_PERF
ggml_graph_compute_perf_stats_node(node, state->shared); ggml_graph_compute_perf_stats_node(node, state->shared);
#endif
} else { } else {
break; break;
} }