vulkan: add barrier after writetimestamp (#21865)

Author: jeffbolznv

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

@@ -13014,6 +13014,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
             if (vk_perf_logger_enabled && vk_perf_logger_concurrent) {
                 ctx->query_node_idx[ctx->query_idx] = node_idx;
                 compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
+                ggml_vk_sync_buffers(ctx, compute_ctx);
             }
         }
         // Add all fused nodes to the unsynchronized lists.
@@ -14503,6 +14504,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         compute_ctx = ggml_vk_get_compute_ctx(ctx);
         ctx->query_idx = 0;
         compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
+        ggml_vk_sync_buffers(ctx, compute_ctx);
     }
 
     ctx->prealloc_y_last_pipeline_used = nullptr;
@@ -14739,6 +14741,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
                 ctx->query_nodes[ctx->query_idx] = cgraph->nodes[i];
                 ctx->query_fusion_names[ctx->query_idx] = fusion_string;
                 compute_ctx->s->buffer->buf.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->query_pool, ctx->query_idx++);
+                ggml_vk_sync_buffers(ctx, compute_ctx);
             } else {
                 // track a fusion string and number of fused ops for the current node_idx
                 ctx->query_fusion_names[i] = fusion_string;