diff --git a/sycl/test/backend/cuda/primary_context.cpp b/sycl/test/backend/cuda/primary_context.cpp
new file mode 100644
index 000000000000..d02b4bca35f6
--- /dev/null
+++ b/sycl/test/backend/cuda/primary_context.cpp
@@ -0,0 +1,116 @@
+// REQUIRES: cuda
+// RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -I%opencl_include_dir -I%cuda_toolkit_include -o %t.out -lcuda -lsycl
+// RUN: env SYCL_DEVICE_TYPE=GPU %t.out
+// NOTE: OpenCL is required for the runtime, even when using the CUDA BE.
+
+//==---------- primary_context.cpp - SYCL cuda primary context test --------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include <CL/sycl.hpp>
+#include <CL/sycl/detail/pi_cuda.hpp>
+#include <cuda.h>
+#include <iostream>
+
+using namespace cl::sycl;
+
+void check(bool condition, const char *conditionString, const char *filename,
+           const long line) noexcept {
+  if (!condition) {
+    std::cerr << "CHECK failed in " << filename << "#" << line << " "
+              << conditionString << "\n";
+    std::abort();
+  }
+}
+
+#define CHECK(CONDITION) check(CONDITION, #CONDITION, __FILE__, __LINE__)
+
+bool isCudaDevice(const device &dev) {
+  const platform platform = dev.get_info<info::device::platform>();
+  const std::string platformVersion =
+      platform.get_info<info::platform::version>();
+  // If using PI_CUDA, don't accept a non-CUDA device
+  return platformVersion.find("CUDA") != std::string::npos;
+}
+
+class cuda_device_selector : public device_selector {
+public:
+  int operator()(const device &dev) const {
+    return isCudaDevice(dev) ? 1 : -1;
+  }
+};
+
+class other_cuda_device_selector : public device_selector {
+public:
+  other_cuda_device_selector(const device &dev) : excludeDevice{dev} {}
+
+  int operator()(const device &dev) const {
+    if (!isCudaDevice(dev)) {
+      return -1;
+    }
+    if (dev.get() == excludeDevice.get()) {
+      // Return only this device if it is the only available
+      return 0;
+    }
+    return 1;
+  }
+
+private:
+  const device &excludeDevice;
+};
+
+int main() {
+  try {
+    context c;
+  } catch (device_error &e) {
+    std::cout << "Failed to create device for context" << std::endl;
+  }
+
+  device DeviceA = cuda_device_selector().select_device();
+  device DeviceB = other_cuda_device_selector(DeviceA).select_device();
+
+  CHECK(isCudaDevice(DeviceA));
+
+  {
+    std::cout << "create single context" << std::endl;
+    context Context(DeviceA, async_handler{}, /*UsePrimaryContext=*/true);
+
+    CUdevice CudaDevice = reinterpret_cast<pi_device>(DeviceA.get())->get();
+    CUcontext CudaContext = reinterpret_cast<pi_context>(Context.get())->get();
+
+    CUcontext PrimaryCudaContext;
+    cuDevicePrimaryCtxRetain(&PrimaryCudaContext, CudaDevice);
+
+    CHECK(CudaContext == PrimaryCudaContext);
+
+    cuDevicePrimaryCtxRelease(CudaDevice);
+  }
+  {
+    std::cout << "create multiple contexts for one device" << std::endl;
+    context ContextA(DeviceA, async_handler{}, /*UsePrimaryContext=*/true);
+    context ContextB(DeviceA, async_handler{}, /*UsePrimaryContext=*/true);
+
+    CUcontext CudaContextA =
+        reinterpret_cast<pi_context>(ContextA.get())->get();
+    CUcontext CudaContextB =
+        reinterpret_cast<pi_context>(ContextB.get())->get();
+
+    CHECK(CudaContextA == CudaContextB);
+  }
+  if (isCudaDevice(DeviceB) && DeviceA.get() != DeviceB.get()) {
+    std::cout << "create multiple contexts for multiple devices" << std::endl;
+    context ContextA(DeviceA, async_handler{}, /*UsePrimaryContext=*/true);
+    context ContextB(DeviceB, async_handler{}, /*UsePrimaryContext=*/true);
+
+    CUcontext CudaContextA =
+        reinterpret_cast<pi_context>(ContextA.get())->get();
+    CUcontext CudaContextB =
+        reinterpret_cast<pi_context>(ContextB.get())->get();
+
+    CHECK(CudaContextA != CudaContextB);
+  }
+}
diff --git a/sycl/test/lit.cfg.py b/sycl/test/lit.cfg.py
index 999b6d1cd558..23cdc5e7e302 100644
--- a/sycl/test/lit.cfg.py
+++ b/sycl/test/lit.cfg.py
@@ -71,6 +71,8 @@
 config.substitutions.append( ('%sycl_include',  config.sycl_include ) )
 config.substitutions.append( ('%opencl_libs_dir',  config.opencl_libs_dir) )
 config.substitutions.append( ('%sycl_source_dir', config.sycl_source_dir) )
+config.substitutions.append( ('%opencl_include_dir',  config.opencl_include_dir) )
+config.substitutions.append( ('%cuda_toolkit_include',  config.cuda_toolkit_include) )
 
 llvm_config.use_clang()
 
diff --git a/sycl/test/lit.site.cfg.py.in b/sycl/test/lit.site.cfg.py.in
index 18ef72f3a94e..4ce4d38cdc91 100644
--- a/sycl/test/lit.site.cfg.py.in
+++ b/sycl/test/lit.site.cfg.py.in
@@ -12,6 +12,8 @@ config.opencl_libs_dir = os.path.dirname("@OpenCL_LIBRARIES@")
 config.sycl_libs_dir = lit_config.params.get('SYCL_LIBS_DIR', "@LLVM_LIBS_DIR@")
 config.target_triple = "@TARGET_TRIPLE@"
 config.host_triple = "@LLVM_HOST_TRIPLE@"
+config.opencl_include_dir = "@OpenCL_INCLUDE_DIR@"
+config.cuda_toolkit_include = "@CUDA_TOOLKIT_INCLUDE@"
 
 config.llvm_enable_projects = "@LLVM_ENABLE_PROJECTS@"