ggml : better PERF prints

diff --git a/src/ggml.c b/src/ggml.c
index 3ee2d0814ec04c07c1c484aa2dd74ff1fced5a8d..f8f73af3e75388ca78ef058da1596f4bb62910e3 100644 (file)
--- a/src/ggml.c
+++ b/src/ggml.c
@@ -11237,9 +11237,9 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
-        perf_total_per_op_us[node->op] += node->perf_time_us;
+        perf_total_per_op_us[node->op] += MAX(1, node->perf_time_us);
 
-        GGML_PRINT(" - %3d: [ %" PRId64 ", %" PRId64 ", %" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
+        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
                 GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
@@ -11253,13 +11253,17 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * node = cgraph->leafs[i];
 
-        GGML_PRINT(" - %3d: [ %" PRId64 ", %" PRId64 "] %8s\n",
+        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n",
                 i,
                 node->ne[0], node->ne[1],
                 GGML_OP_LABEL[node->op]);
     }
 
     for (int i = 0; i < GGML_OP_COUNT; i++) {
+        if (perf_total_per_op_us[i] == 0) {
+            continue;
+        }
+
         GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_LABEL[i], (double) perf_total_per_op_us[i] / 1000.0);
     }