vulkan: Submit once enough matmul work has been recorded

I've been seeing significantly worse performance for tg with flash attention
enabled vs disabled, and it seems to be related to the submit heuristic.
Change the heuristic to check how many bytes worth of weight matrix are
used and flush every 100MB. This seems to resolve the issue, and also
increases perf for non-FA a bit.

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -8267,17 +8267,26 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     bool first_node_in_batch = true; // true if next node will be first node in a batch
     int submit_node_idx = 0; // index to first node in a batch
 
-    // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution.
-    // Start with a smaller count to get work submitted right away, and increase it after each submit.
-    int nodes_per_submit = 20;
+    // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
+    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB.
+    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
+    int nodes_per_submit = 100;
     int submitted_nodes = 0;
     int submit_count = 0;
+    uint64_t mul_mat_bytes = 0;
+    uint64_t mul_mat_bytes_per_submit = 100*1000*1000;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
-        bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
+
+        bool submit = (submitted_nodes >= nodes_per_submit) ||
+                      (mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+                      (i == last_node);
 
         bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
 
@@ -8294,14 +8303,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         if (submit) {
             first_node_in_batch = true;
             submitted_nodes = 0;
-            switch (submit_count) {
-            case 0:
-                nodes_per_submit = 50;
-                break;
-            default:
-                nodes_per_submit = 100;
-                break;
-            }
+            mul_mat_bytes = 0;
             submit_count++;
         }
     }