vulkan: get the first command buffer submitted sooner

This is an incremental improvement over #9118 to get work to the GPU a bit
sooner. The first part is to start with a smaller number of nodes before
the first submit, and ramp it up to the current 100 nodes/submit. The
second part is to reduce the dryrun overhead for all the nodes that just
need to request descriptor space.

With these changes I get around 1-2% speedup on RTX 4070 combined with my
old Haswell-era CPU.

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

@@ -5630,6 +5630,48 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         } else {
             compute_ctx = ctx->compute_ctx.lock();
         }
+    } else {
+        switch (node->op) {
+        case GGML_OP_REPEAT:
+        case GGML_OP_ACC:
+        case GGML_OP_GET_ROWS:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+        case GGML_OP_CLAMP:
+        case GGML_OP_PAD:
+        case GGML_OP_CPY:
+        case GGML_OP_CONT:
+        case GGML_OP_DUP:
+        case GGML_OP_NORM:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_UNARY:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_IM2COL:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_POOL_2D:
+        case GGML_OP_LEAKY_RELU:
+            {
+                // These operations all go through ggml_vk_op_f32, so short-circuit and
+                // do the only thing needed for the dryrun.
+                vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
+                ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+                return false;
+            }
+        default:
+            break;
+        }
     }
 
     switch (node->op) {
@@ -6359,16 +6401,17 @@ 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 submit_count node to overlap CPU cmdbuffer generation with GPU execution
-    constexpr int submit_count = 100;
+    // 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;
     int submitted_nodes = 0;
+    int submit_count = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
-        bool submit = (submitted_nodes >= submit_count) || (i == last_node);
-
+        bool submit = (submitted_nodes >= nodes_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);
 
@@ -6385,6 +6428,15 @@ 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;
+            }
+            submit_count++;
         }
     }