Skip barriers of noops

picked from ik_llama.cpp, a llama_cpp fork maintained by Iwan Kawrakow

Co-Authored-By: Kawrakow <[email protected]>

Author: Nexesenex

ggml/include/ggml.h

@@ -734,6 +734,12 @@ extern "C" {
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
+
+    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+
+    GGML_API bool ggml_is_noop(const struct ggml_tensor * tensor);
+
     GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
     GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 

ggml/src/ggml.c

@@ -3773,6 +3773,12 @@ bool ggml_is_empty(const struct ggml_tensor * tensor) {
     return false;
 }
 
+bool ggml_is_noop(const struct ggml_tensor * tensor) {
+    return tensor->op == GGML_OP_NONE || tensor->op == GGML_OP_RESHAPE ||
+           tensor->op == GGML_OP_VIEW || tensor->op == GGML_OP_PERMUTE || tensor->op == GGML_OP_TRANSPOSE ||
+           ggml_is_empty(tensor) ? true : false;
+}
+
 bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -20035,6 +20041,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
+        if (ggml_is_noop(node)) continue;
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&