diff --git a/README.md b/README.md
index a2e9df0..7098518 100644
--- a/README.md
+++ b/README.md
@@ -80,4 +80,3 @@ For faster parallel builds, you can instead generate [ninja](https://ninja-build
 Finish porting
 - [ ] shader hot reloading
 - [ ] descriptor gen
-- [ ] acceleration structures
diff --git a/rvk/device.cpp b/rvk/device.cpp
index 80ce5f7..519b143 100644
--- a/rvk/device.cpp
+++ b/rvk/device.cpp
@@ -607,11 +607,11 @@ u32 Device::queue_index(Queue_Family family) {
 
 void Device::submit(Commands& cmds, u32 index, Fence& fence) {
 
-    VkCommandBuffer vk_cmds = cmds;
+    cmds.end();
 
     VkCommandBufferSubmitInfo cmd_info = {};
     cmd_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_SUBMIT_INFO;
-    cmd_info.commandBuffer = vk_cmds;
+    cmd_info.commandBuffer = cmds;
 
     VkSubmitInfo2 submit_info = {};
     submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2;
@@ -627,6 +627,9 @@ void Device::submit(Commands& cmds, u32 index, Fence& fence) {
 
 void Device::submit(Commands& cmds, u32 index, Slice<Sem_Ref> signal, Slice<Sem_Ref> wait,
                     Fence& fence) {
+
+    cmds.end();
+
     Region(R) {
 
         Vec<VkSemaphoreSubmitInfo, Mregion<R>> vk_signal(signal.length());
diff --git a/rvk/rvk.cpp b/rvk/rvk.cpp
index 076c954..a099b05 100644
--- a/rvk/rvk.cpp
+++ b/rvk/rvk.cpp
@@ -439,7 +439,6 @@ void Vk::end_frame(Image_View& output) {
         frame.cmds.reset();
         compositor->render(frame.cmds, state.frame_index, state.swapchain_index, state.has_imgui,
                            output);
-        frame.cmds.end();
 
         // Wait for frame available before running the submit; signal frame complete on finish
         frame.wait_for.push(