[ET-VK] 2/n Split dispatches between multiple command buffers. Add semaphore support to Adapter::submit_cmd. (#12522)

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at
bottom):
* #12525
* #12523
* __->__ #12522
* #12519

This diff adds semaphore support to `Adapter::submit_cmd` by modifying
its function signature to include `wait_semaphore` and
`signal_semaphore` parameters. The updated function now takes into
account the new semaphore parameters and correctly configures the
pipeline stages using `VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT`.

Differential Revision:
[D78360041](https://our.internmc.facebook.com/intern/diff/D78360041/)

---------

Co-authored-by: Stephen Jia <[email protected]>

Author: trivedivivek

backends/vulkan/runtime/vk_api/Adapter.cpp

@@ -307,17 +307,22 @@ void Adapter::return_queue(Adapter::Queue& compute_queue) {
 void Adapter::submit_cmd(
     const Adapter::Queue& device_queue,
     VkCommandBuffer cmd,
-    VkFence fence) {
+    VkFence fence,
+    VkSemaphore wait_semaphore,
+    VkSemaphore signal_semaphore) {
+  const VkPipelineStageFlags flags = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+  const bool set_wait_semaphore = wait_semaphore != VK_NULL_HANDLE;
+  const bool set_signal_semaphore = signal_semaphore != VK_NULL_HANDLE;
   const VkSubmitInfo submit_info{
       VK_STRUCTURE_TYPE_SUBMIT_INFO, // sType
       nullptr, // pNext
-      0u, // waitSemaphoreCount
-      nullptr, // pWaitSemaphores
-      nullptr, // pWaitDstStageMask
+      set_wait_semaphore ? 1u : 0u, // waitSemaphoreCount
+      set_wait_semaphore ? &wait_semaphore : nullptr, // pWaitSemaphores
+      &flags, // pWaitDstStageMask
       1u, // commandBufferCount
       &cmd, // pCommandBuffers
-      0u, // signalSemaphoreCount
-      nullptr, // pSignalSemaphores
+      set_signal_semaphore ? 1u : 0u, // signalSemaphoreCount
+      set_signal_semaphore ? &signal_semaphore : nullptr, // pSignalSemaphores
   };
 
   std::lock_guard<std::mutex> queue_lock(

backends/vulkan/runtime/vk_api/Adapter.h

@@ -242,8 +242,12 @@ class Adapter final {
 
   // Command Buffer Submission
 
-  void
-  submit_cmd(const Queue&, VkCommandBuffer, VkFence fence = VK_NULL_HANDLE);
+  void submit_cmd(
+      const Queue&,
+      VkCommandBuffer,
+      VkFence fence = VK_NULL_HANDLE,
+      VkSemaphore wait_semaphore = VK_NULL_HANDLE,
+      VkSemaphore signal_semaphore = VK_NULL_HANDLE);
 
   std::string stringize() const;
   friend std::ostream& operator<<(std::ostream&, const Adapter&);