[SYCL][Graph]Fix and add E2E tests for using local accessors in graphs

sycl/cmake/modules/UnifiedRuntimeTag.cmake

@@ -1,7 +1,7 @@
-# commit f01741af022cfe82afcb026b9aa0be251eb6a497
-# Merge: 004d2474 85bb5f62
-# Author: Callum Fare <callum@codeplay.com>
-# Date:   Tue Nov 5 13:39:53 2024 +0000
-#     Merge pull request #2260 from nrspruit/refactor_l0_default_init
-#     [L0] Refactor to remove default constructor inits
-set(UNIFIED_RUNTIME_TAG f01741af022cfe82afcb026b9aa0be251eb6a497)
+# commit 3edf99755ce2af3b53102a7d8438e0fe969efac3
+# Merge: 5955bad3 0b968661
+# Author: Ross Brunton <ross@codeplay.com>
+# Date:   Wed Nov 6 11:07:29 2024 +0000
+#    Merge pull request #2082 from RossBrunton/ross/multiadapt
+#    [CI] Add "loader" support to conformance testing
+set(UNIFIED_RUNTIME_TAG 3edf99755ce2af3b53102a7d8438e0fe969efac3)

sycl/test-e2e/Graph/Explicit/local_accessor.cpp

@@ -0,0 +1,10 @@
+// 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 %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+// Extra run to check for immediate-command-list in Level Zero
+// RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+
+#define GRAPH_E2E_EXPLICIT
+
+#include "../Inputs/local_accessor.cpp"

sycl/test-e2e/Graph/Inputs/local_accessor.cpp

@@ -0,0 +1,54 @@
+// Tests basic adding of nodes with local accessors,
+// and submission of the graph.
+
+#include "../graph_common.hpp"
+
+int main() {
+  queue Queue{};
+
+  using T = int;
+
+  const size_t LocalSize = 128;
+
+  std::vector<T> DataA(Size), DataB(Size), DataC(Size);
+
+  std::iota(DataA.begin(), DataA.end(), 10);
+
+  std::vector<T> ReferenceA(DataA);
+
+  exp_ext::command_graph Graph{Queue.get_context(), Queue.get_device()};
+
+  T *PtrA = malloc_device<T>(Size, Queue);
+
+  Queue.copy(DataA.data(), PtrA, Size);
+  Queue.wait_and_throw();
+
+  auto node = add_node(Graph, Queue, [&](handler &CGH) {
+    local_accessor<T, 1> LocalMem(LocalSize, CGH);
+
+    CGH.parallel_for(nd_range({Size}, {LocalSize}), [=](nd_item<1> Item) {
+      LocalMem[Item.get_local_linear_id()] = Item.get_global_linear_id() * 2;
+      PtrA[Item.get_global_linear_id()] += LocalMem[Item.get_local_linear_id()];
+    });
+  });
+
+  auto GraphExec = Graph.finalize();
+
+  for (unsigned n = 0; n < Iterations; n++) {
+    Queue.submit([&](handler &CGH) { CGH.ext_oneapi_graph(GraphExec); });
+  }
+
+  Queue.wait_and_throw();
+
+  Queue.copy(PtrA, DataA.data(), Size);
+  Queue.wait_and_throw();
+
+  free(PtrA, Queue);
+
+  for (size_t i = 0; i < Size; i++) {
+    T Ref = 10 + i + (i * 2);
+    check_value(i, Ref, ReferenceA[i], "PtrA");
+  }
+
+  return 0;
+}

sycl/test-e2e/Graph/Inputs/whole_update_local_acc.cpp

@@ -0,0 +1,93 @@
+// Tests whole graph update of nodes with local accessors,
+// and submission of the graph.
+
+#include "../graph_common.hpp"
+
+using T = int;
+
+auto add_graph_node(
+    exp_ext::command_graph<exp_ext::graph_state::modifiable> &Graph,
+    queue &Queue, size_t Size, size_t LocalSize, T *Ptr) {
+  return add_node(Graph, Queue, [&](handler &CGH) {
+    local_accessor<T, 1> LocalMem(LocalSize, CGH);
+
+    CGH.parallel_for(nd_range({Size}, {LocalSize}), [=](nd_item<1> Item) {
+      LocalMem[Item.get_local_linear_id()] = Item.get_global_linear_id() * 2;
+      Ptr[Item.get_global_linear_id()] +=
+          LocalMem[Item.get_local_linear_id()] + Item.get_local_range(0);
+    });
+  });
+}
+int main() {
+  queue Queue{};
+
+  const size_t LocalSize = 128;
+
+  std::vector<T> DataA(Size), DataB(Size);
+
+  std::iota(DataA.begin(), DataA.end(), 10);
+  std::iota(DataB.begin(), DataB.end(), 10);
+
+  std::vector<T> ReferenceA(DataA), ReferenceB(DataB);
+
+  exp_ext::command_graph GraphA{Queue.get_context(), Queue.get_device()};
+
+  T *PtrA = malloc_device<T>(Size, Queue);
+  T *PtrB = malloc_device<T>(Size, Queue);
+
+  Queue.copy(DataA.data(), PtrA, Size);
+  Queue.copy(DataB.data(), PtrB, Size);
+  Queue.wait_and_throw();
+
+  auto NodeA = add_graph_node(GraphA, Queue, Size, LocalSize / 2, PtrA);
+
+  auto GraphExecA = GraphA.finalize(exp_ext::property::graph::updatable{});
+
+  // Create second graph for whole graph update with a different local size
+  exp_ext::command_graph GraphB{Queue.get_context(), Queue.get_device()};
+  auto NodeB = add_graph_node(GraphB, Queue, Size, LocalSize, PtrB);
+
+  // Execute graphs before updating and check outputs
+  for (unsigned n = 0; n < Iterations; n++) {
+    Queue.submit([&](handler &CGH) { CGH.ext_oneapi_graph(GraphExecA); });
+  }
+
+  Queue.wait_and_throw();
+
+  Queue.copy(PtrA, DataA.data(), Size);
+  Queue.copy(PtrB, DataB.data(), Size);
+  Queue.wait_and_throw();
+
+  for (size_t i = 0; i < Size; i++) {
+    T RefA = 10 + i + (i * 2) + LocalSize / 2;
+    T RefB = 10 + i;
+    check_value(i, RefA, ReferenceA[i], "PtrA");
+    check_value(i, RefB, ReferenceB[i], "PtrB");
+  }
+
+  // Update GraphExecA using whole graph update
+
+  GraphExecA.update(GraphB);
+
+  // Execute graphs again and check outputs
+  for (unsigned n = 0; n < Iterations; n++) {
+    Queue.submit([&](handler &CGH) { CGH.ext_oneapi_graph(GraphExecA); });
+  }
+
+  Queue.wait_and_throw();
+
+  Queue.copy(PtrA, DataA.data(), Size);
+  Queue.copy(PtrB, DataB.data(), Size);
+  Queue.wait_and_throw();
+
+  for (size_t i = 0; i < Size; i++) {
+    T RefA = 10 + i + (i * 2) + LocalSize / 2;
+    T RefB = 10 + i + (i * 2) + LocalSize;
+    check_value(i, RefA, ReferenceA[i], "PtrA");
+    check_value(i, RefB, ReferenceB[i], "PtrB");
+  }
+
+  free(PtrA, Queue);
+  free(PtrB, Queue);
+  return 0;
+}

sycl/test-e2e/Graph/RecordReplay/local_accessor.cpp

@@ -0,0 +1,10 @@
+// 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 %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+// Extra run to check for immediate-command-list in Level Zero
+// RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+
+#define GRAPH_E2E_RECORD_REPLAY
+
+#include "../Inputs/local_accessor.cpp"

sycl/test-e2e/Graph/Update/Explicit/whole_update_local_acc.cpp

@@ -0,0 +1,10 @@
+// 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 %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+// Extra run to check for immediate-command-list in Level Zero
+// RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+
+#define GRAPH_E2E_EXPLICIT
+
+#include "../../Inputs/whole_update_local_acc.cpp"

sycl/test-e2e/Graph/Update/RecordReplay/whole_update_local_acc.cpp

@@ -0,0 +1,10 @@
+// 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 %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+// Extra run to check for immediate-command-list in Level Zero
+// RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
+
+#define GRAPH_E2E_RECORD_REPLAY
+
+#include "../../Inputs/whole_update_local_acc.cpp"