From 8cc68177e7c418b992dd1d8b54c4c142000db513 Mon Sep 17 00:00:00 2001
From: Matthew Michel <matthew.michel@intel.com>
Date: Fri, 23 Oct 2026 14:23:40 -0700
Subject: [PATCH] [SYCL][Graph][Test] Add native recording test that
 instantiates the graph multiple times

---
 .../NativeRecording/multiple_exec_graphs.cpp  | 101 ++++++++++++++++++
 1 file changed, 101 insertions(+)
 create mode 100644 sycl/test-e2e/Graph/RecordReplay/NativeRecording/multiple_exec_graphs.cpp

diff --git a/sycl/test-e2e/Graph/RecordReplay/NativeRecording/multiple_exec_graphs.cpp b/sycl/test-e2e/Graph/RecordReplay/NativeRecording/multiple_exec_graphs.cpp
new file mode 100644
index 0000000000000..540bd30d06649
--- /dev/null
+++ b/sycl/test-e2e/Graph/RecordReplay/NativeRecording/multiple_exec_graphs.cpp
@@ -0,0 +1,101 @@
+// REQUIRES: level_zero_v2_adapter && (arch-intel_gpu_bmg_g21 || arch-intel_gpu_bmg_g31 || arch-intel_gpu_cri)
+
+// RUN: %{build} -o %t.out
+// RUN: %{run} %t.out
+// Extra run to check for leaks in Level Zero using UR_L0_LEAKS_DEBUG
+// RUN: %if level_zero %{%{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+
+// Submits several executable graphs finalized from one native recording graph
+// round-robin across queues so they may run concurrently. All instances share
+// the same USM, so the graph only reads inputs and accumulates with atomics.
+
+#include "../../graph_common.hpp"
+
+#include <sycl/atomic_ref.hpp>
+#include <sycl/properties/all_properties.hpp>
+
+template <typename T>
+using device_atomic_ref =
+    atomic_ref<T, memory_order::relaxed, memory_scope::device,
+               access::address_space::global_space>;
+
+int main() {
+  device Dev;
+  context Ctx{Dev};
+
+  constexpr size_t NumQueues = 4;
+  constexpr size_t NumExecGraphs = 2 * NumQueues;
+  std::vector<queue> Queues;
+  for (size_t i = 0; i < NumQueues; i++) {
+    Queues.emplace_back(Ctx, Dev, property::queue::in_order{});
+  }
+
+  exp_ext::command_graph Graph{
+      Ctx, Dev, {exp_ext::property::graph::enable_native_recording{}}};
+
+  int *InputA = malloc_device<int>(Size, Dev, Ctx);
+  int *InputB = malloc_device<int>(Size, Dev, Ctx);
+  int *Products = malloc_device<int>(Size, Dev, Ctx);
+  int *Sum = malloc_device<int>(1, Dev, Ctx);
+
+  std::vector<int> HostA(Size), HostB(Size);
+  std::iota(HostA.begin(), HostA.end(), 0);
+  std::iota(HostB.begin(), HostB.end(), 1);
+
+  queue &RecordQueue = Queues[0];
+  RecordQueue.copy(HostA.data(), InputA, Size);
+  RecordQueue.copy(HostB.data(), InputB, Size);
+  RecordQueue.fill(Products, 0, Size);
+  RecordQueue.fill(Sum, 0, 1);
+  RecordQueue.wait_and_throw();
+
+  Graph.begin_recording(RecordQueue);
+
+  RecordQueue.parallel_for(range<1>{Size}, [=](id<1> idx) {
+    device_atomic_ref<int>{Products[idx]} += InputA[idx] * InputB[idx];
+  });
+
+  RecordQueue.parallel_for(range<1>{Size}, [=](id<1> idx) {
+    device_atomic_ref<int>{Sum[0]} += InputA[idx];
+  });
+
+  Graph.end_recording(RecordQueue);
+
+  std::vector<exp_ext::command_graph<exp_ext::graph_state::executable>>
+      ExecGraphs;
+  for (size_t i = 0; i < NumExecGraphs; i++) {
+    ExecGraphs.push_back(Graph.finalize());
+  }
+
+  for (unsigned n = 0; n < Iterations; n++) {
+    for (size_t i = 0; i < NumExecGraphs; i++) {
+      Queues[i % NumQueues].ext_oneapi_graph(ExecGraphs[i]);
+    }
+  }
+
+  for (queue &Q : Queues) {
+    Q.wait_and_throw();
+  }
+
+  std::vector<int> HostProducts(Size);
+  int HostSum = 0;
+  RecordQueue.copy(Products, HostProducts.data(), Size);
+  RecordQueue.copy(Sum, &HostSum, 1);
+  RecordQueue.wait_and_throw();
+
+  const int TotalExecutions = NumExecGraphs * Iterations;
+  int ExpectedSum = 0;
+  for (size_t i = 0; i < Size; i++) {
+    int Expected = TotalExecutions * HostA[i] * HostB[i];
+    assert(check_value(i, Expected, HostProducts[i], "Products"));
+    ExpectedSum += TotalExecutions * HostA[i];
+  }
+  assert(check_value(0, ExpectedSum, HostSum, "Sum"));
+
+  free(InputA, Ctx);
+  free(InputB, Ctx);
+  free(Products, Ctx);
+  free(Sum, Ctx);
+
+  return 0;
+}