[ET-VK] 5/n Split dispatches between multiple command buffers. Track previously submitted command buffers in context and add function to execute all previous command buffers.

backends/vulkan/runtime/api/Context.cpp

@@ -217,7 +217,7 @@ void Context::submit_cmd_to_gpu(VkFence fence_handle, const bool final_use) {
 }
 
 void Context::flush() {
-  VK_CHECK(vkQueueWaitIdle(queue()));
+  VK_CHECK(vkQueueWaitIdle(queue().handle));
 
   command_pool_.flush();
   descriptor_pool_.flush();

backends/vulkan/runtime/api/Context.h

@@ -90,8 +90,8 @@ class Context final {
     return device_;
   }
 
-  inline VkQueue queue() {
-    return queue_.handle;
+  inline vkapi::Adapter::Queue& queue() {
+    return queue_;
   }
 
   // Device Caches
@@ -230,6 +230,10 @@ class Context final {
       VkFence fence_handle = VK_NULL_HANDLE,
       const bool final_use = false);
 
+  vkapi::CommandBuffer& extract_cmd() {
+    return cmd_;
+  }
+
   void flush();
 
 #ifdef VULKAN_DEBUG

backends/vulkan/runtime/graph/ComputeGraph.cpp

@@ -158,6 +158,7 @@ ComputeGraph::~ComputeGraph() {
 
   prepack_nodes_.clear();
   execute_nodes_.clear();
+  deferred_cmd_list_.clear();
 
   context_->flush();
 }
@@ -767,6 +768,30 @@ void ComputeGraph::submit_current_cmd_and_wait(const bool final_use) {
   context_->fences().return_fence(fence);
 }
 
+void ComputeGraph::submit_deferred_cmds() {
+  VkSemaphore prev_semaphore = VK_NULL_HANDLE;
+  vkapi::VulkanFence fence = context_->fences().get_fence();
+
+  for (uint32_t i = 0; i < deferred_cmd_list_.size(); i++) {
+    auto& cmd = deferred_cmd_list_[i];
+    VkSemaphore wait_semaphore = prev_semaphore;
+    VkSemaphore signal_semaphore = cmd.get_signal_semaphore();
+    prev_semaphore = signal_semaphore;
+
+    if (cmd) {
+      cmd.end();
+      context_->adapter_ptr()->submit_cmd(
+          context_->queue(),
+          cmd.get_submit_handle(false),
+          i == (deferred_cmd_list_.size() - 1) ? fence.get_submit_handle() : VK_NULL_HANDLE,
+          wait_semaphore,
+          signal_semaphore);
+    }
+  }
+  fence.wait();
+  context_->fences().return_fence(fence);
+}
+
 void ComputeGraph::prepack() {
   int i = 0;
   bool submitted = false;
@@ -805,6 +830,7 @@ void ComputeGraph::prepack() {
 }
 
 void ComputeGraph::encode_execute() {
+  deferred_cmd_list_.clear();
   context_->flush();
   context_->set_cmd(/*reusable = */ true);
 
@@ -813,13 +839,12 @@ void ComputeGraph::encode_execute() {
   for (std::unique_ptr<ExecuteNode>& node : execute_nodes_) {
     node->encode(this);
   }
+
+  deferred_cmd_list_.emplace_back(std::move(context_->extract_cmd()));
 }
 
 void ComputeGraph::execute() {
-  vkapi::VulkanFence fence = context_->fences().get_fence();
-  context_->submit_cmd_to_gpu(fence.get_submit_handle());
-  fence.wait();
-  context_->fences().return_fence(fence);
+  submit_deferred_cmds();
   execute_count_++;
 }
 

backends/vulkan/runtime/graph/ComputeGraph.h

@@ -193,6 +193,9 @@ class ComputeGraph final {
   // Utility constexpr to express byte quantities
   constexpr static size_t MB = 1024 * 1024;
 
+  // List of command buffers deferred for submission
+  std::vector<vkapi::CommandBuffer> deferred_cmd_list_;
+
  protected:
   size_t values_in_use_ = 0;
   size_t execute_count_ = 0;
@@ -851,6 +854,11 @@ class ComputeGraph final {
    */
   void submit_current_cmd_and_wait(const bool final_use = false);
 
+  /*
+   * Submits all the commands gathered in deferred_cmd_bufs_ to the GPU.
+   */
+  void submit_deferred_cmds();
+
  public:
   //
   // Graph Prepacking