From 71a594ae901ec6cd16143678f0748e893cae050e Mon Sep 17 00:00:00 2001
From: Guo Yejun <yejun.guo@intel.com>
Date: Tue, 20 Feb 2024 03:42:03 -0800
Subject: [PATCH] [SYCL][Graph] submit.cpp: check test result according to spec

---
 sycl/test-e2e/Graph/Threading/submit.cpp | 41 ++++++++++++++----------
 1 file changed, 24 insertions(+), 17 deletions(-)

diff --git a/sycl/test-e2e/Graph/Threading/submit.cpp b/sycl/test-e2e/Graph/Threading/submit.cpp
index 1a815b188630e..5029647b5bdcb 100644
--- a/sycl/test-e2e/Graph/Threading/submit.cpp
+++ b/sycl/test-e2e/Graph/Threading/submit.cpp
@@ -4,11 +4,11 @@
 // Extra run to check for immediate-command-list in Level Zero
 // RUN: %if level_zero && linux %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
 
-// Test submitting a graph in a threaded situation.
-// Note that we do not check the outputs because multiple concurrent executions
-// is indeterministic (and depends on the backend command management).
-// However, this test verifies that concurrent graph submissions do not trigger
-// errors nor memory leaks.
+// Test submitting a graph multiple times in a threaded situation.
+// According to spec: If graph is submitted multiple times, dependencies are
+// automatically added by the runtime to prevent concurrent executions of an
+// identical graph, and so the result is deterministic and we can check the
+// results.
 
 #include "../graph_common.hpp"
 
@@ -20,6 +20,7 @@ int main() {
   using T = int;
 
   const unsigned NumThreads = std::thread::hardware_concurrency();
+  const unsigned SubmitsPerThread = 128;
   std::vector<T> DataA(Size), DataB(Size), DataC(Size);
 
   std::iota(DataA.begin(), DataA.end(), 1);
@@ -27,8 +28,8 @@ int main() {
   std::iota(DataC.begin(), DataC.end(), 1000);
 
   std::vector<T> ReferenceA(DataA), ReferenceB(DataB), ReferenceC(DataC);
-  calculate_reference_data(NumThreads, Size, ReferenceA, ReferenceB,
-                           ReferenceC);
+  calculate_reference_data(NumThreads * SubmitsPerThread, Size, ReferenceA,
+                           ReferenceB, ReferenceC);
 
   exp_ext::command_graph Graph{Queue.get_context(), Queue.get_device()};
 
@@ -45,34 +46,40 @@ int main() {
   run_kernels_usm(Queue, Size, PtrA, PtrB, PtrC);
   Graph.end_recording();
 
-  std::vector<exp_ext::command_graph<exp_ext::graph_state::executable>>
-      GraphExecs;
-  for (unsigned i = 0; i < NumThreads; ++i) {
-    GraphExecs.push_back(Graph.finalize());
-  }
+  auto GraphExec = Graph.finalize();
 
   Barrier SyncPoint{NumThreads};
 
-  auto SubmitGraph = [&](int ThreadNum) {
+  auto SubmitGraph = [&]() {
     SyncPoint.wait();
-    Queue.submit([&](sycl::handler &CGH) {
-      CGH.ext_oneapi_graph(GraphExecs[ThreadNum]);
-    });
+    for (unsigned i = 0; i < SubmitsPerThread; ++i) {
+      Queue.submit(
+          [&](sycl::handler &CGH) { CGH.ext_oneapi_graph(GraphExec); });
+    }
   };
 
   std::vector<std::thread> Threads;
   Threads.reserve(NumThreads);
 
   for (unsigned i = 0; i < NumThreads; ++i) {
-    Threads.emplace_back(SubmitGraph, i);
+    Threads.emplace_back(SubmitGraph);
   }
 
   for (unsigned i = 0; i < NumThreads; ++i) {
     Threads[i].join();
   }
 
+  Queue.copy(PtrA, DataA.data(), Size);
+  Queue.copy(PtrB, DataB.data(), Size);
+  Queue.copy(PtrC, DataC.data(), Size);
   Queue.wait_and_throw();
 
+  for (int i = 0; i < Size; ++i) {
+    check_value(i, ReferenceA[i], DataA[i], "A");
+    check_value(i, ReferenceB[i], DataB[i], "B");
+    check_value(i, ReferenceC[i], DataC[i], "C");
+  }
+
   free(PtrA, Queue);
   free(PtrB, Queue);
   free(PtrC, Queue);