diff --git a/SYCL/Basic/alloc_pinned_host_memory.cpp b/SYCL/Basic/alloc_pinned_host_memory.cpp
index 0d8a026750..16e64c3eb8 100644
--- a/SYCL/Basic/alloc_pinned_host_memory.cpp
+++ b/SYCL/Basic/alloc_pinned_host_memory.cpp
@@ -39,4 +39,5 @@ int main() {
 // CHECK:---> piMemBufferCreate
 // CHECK:---> piMemBufferCreate
 // CHECK-NEXT: {{.*}} : {{.*}}
+// CHECK-NEXT: {{.*}} : {{.*}}
 // CHECK-NEXT: {{.*}} : 17
diff --git a/SYCL/Basic/buffer/native_buffer_creation_flags.cpp b/SYCL/Basic/buffer/native_buffer_creation_flags.cpp
index bb29cef6bc..3b2655cf3d 100644
--- a/SYCL/Basic/buffer/native_buffer_creation_flags.cpp
+++ b/SYCL/Basic/buffer/native_buffer_creation_flags.cpp
@@ -21,6 +21,7 @@ int main() {
     // buffer is created with the PI_MEM_FLAGS_HOST_PTR_USE flag.
     // CHECK: piMemBufferCreate
     // CHECK-NEXT: {{.*}} : {{.*}}
+    // CHECK-NEXT: {{.*}} : {{.*}}
     // CHECK-NEXT: {{.*}} : 9
     auto BufAcc = Buf.get_access<access::mode::read>(Cgh);
     Cgh.single_task<Foo>([=]() { int A = BufAcc[0]; });
diff --git a/SYCL/Basic/multi_device_context.cpp b/SYCL/Basic/multi_device_context.cpp
new file mode 100755
index 0000000000..351b87dcfd
--- /dev/null
+++ b/SYCL/Basic/multi_device_context.cpp
@@ -0,0 +1,64 @@
+// RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
+// RUN: %CPU_RUN_PLACEHOLDER %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+
+// REQUIRES: cuda
+
+#include <stdlib.h>
+#include <sycl/sycl.hpp>
+
+sycl::event add(sycl::queue &q, sycl::buffer<int> &buff, int *usm,
+                sycl::event &e) {
+  return q.submit([&](sycl::handler &cgh) {
+    auto acc = buff.get_access<sycl::access::mode::read_write>(cgh);
+    cgh.depends_on(e);
+    cgh.single_task([=]() { acc[0] += *usm; });
+  });
+}
+
+int main() {
+  sycl::platform plat = sycl::platform::get_platforms()[0];
+  auto devices = plat.get_devices();
+  if (devices.size() < 2) {
+    std::cout << "Need two devices for the test!" << std::endl;
+    return 0;
+  }
+
+  sycl::device dev1 = devices[0];
+  sycl::device dev2 = devices[1];
+
+  sycl::context ctx{{dev1, dev2}};
+
+  sycl::queue q1{ctx, dev1};
+  sycl::queue q2{ctx, dev2};
+
+  int a = 1;
+  int b = 2;
+  {
+    sycl::buffer<int> buff1(&a, 1);
+    sycl::buffer<int> buff2(&b, 1);
+
+    // Test copying usm.
+    int *usm1 = sycl::malloc_device<int>(1, q1);
+    int *usm2 = sycl::malloc_device<int>(1, q2);
+    sycl::event e1 = q1.fill(usm1, 4, 1);
+    sycl::event e2 = q2.fill(usm2, 5, 1);
+
+    // Test combination of usm and buffers in a kernel.
+    sycl::event e3 = add(q1, buff1, usm1, e1);
+    sycl::event e4 = add(q2, buff2, usm2, e2);
+
+    // Change values in usm to ensure results are distinct.
+    sycl::event e5 = q1.fill(usm1, 5, 1, e3);
+    sycl::event e6 = q2.fill(usm2, 4, 1, e4);
+
+    // Use each buffer on the other device than before - tests that copying
+    // between devices works.
+    add(q1, buff2, usm1, e5);
+    add(q2, buff1, usm2, e6);
+  }
+  assert(a == 1 + 2 * 4);
+  assert(b == 2 + 2 * 5);
+
+  return 0;
+}
diff --git a/SYCL/Basic/use_pinned_host_memory.cpp b/SYCL/Basic/use_pinned_host_memory.cpp
index 9efb9115b9..18378945ed 100644
--- a/SYCL/Basic/use_pinned_host_memory.cpp
+++ b/SYCL/Basic/use_pinned_host_memory.cpp
@@ -43,4 +43,5 @@ int main() {
 
 // CHECK:---> piMemBufferCreate
 // CHECK-NEXT: {{.*}} : {{.*}}
+// CHECK-NEXT: {{.*}} : {{.*}}
 // CHECK-NEXT: {{.*}} : 17
diff --git a/SYCL/Tracing/pi_tracing_test.cpp b/SYCL/Tracing/pi_tracing_test.cpp
index c91e93cd21..869a612921 100644
--- a/SYCL/Tracing/pi_tracing_test.cpp
+++ b/SYCL/Tracing/pi_tracing_test.cpp
@@ -13,6 +13,7 @@
 // CHECK:  <nullptr>
 // CHECK: ---> piMemBufferCreate(
 // CHECK-NEXT:  <unknown> : {{0[xX]?[0-9a-fA-F]*}}
+// CHECK-NEXT:  <unknown> : {{0[xX]?[0-9a-fA-F]*}}
 // CHECK-NEXT:  <unknown> : 1
 // CHECK-NEXT:  <unknown> : 40
 // CHECK-NEXT:  <unknown> : 0