rough impl for perf testing.

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

sycl/include/sycl/detail/pi.def

@@ -140,5 +140,7 @@ _PI_API(piPluginGetLastError)
 
 _PI_API(piTearDown)
 
+_PI_API(piextEnablePeer)
+
 
 #undef _PI_API

sycl/include/sycl/detail/pi.h

@@ -981,6 +981,8 @@ __SYCL_EXPORT pi_result piDevicesGet(pi_platform platform,
                                      pi_uint32 num_entries, pi_device *devices,
                                      pi_uint32 *num_devices);
 
+__SYCL_EXPORT pi_result piextEnablePeer(pi_queue command_queue, pi_queue peer_queue);
+
 /// Returns requested info for provided native device
 /// Return PI_DEVICE_INFO_EXTENSION_DEVICELIB_ASSERT for
 /// PI_DEVICE_INFO_EXTENSIONS query when the device supports native asserts

sycl/include/sycl/queue.hpp

@@ -62,6 +62,15 @@ class queue;
 template <backend BackendName, class SyclObjectT>
 auto get_native(const SyclObjectT &Obj)
     -> backend_return_t<BackendName, SyclObjectT>;
+namespace ext {
+namespace oneapi {
+namespace experimental {
+
+bool __SYCL_EXPORT ext_oneapi_enable_peer_access(const queue &active, const queue &peer);
+
+} // namespace experimental
+} // namespace oneapi
+} // namespace ext
 
 namespace detail {
 class queue_impl;
@@ -85,6 +94,9 @@ static event submitAssertCapture(queue &, event &, queue *,
 /// \ingroup sycl_api
 class __SYCL_EXPORT queue {
 public:
+
+friend bool ext::oneapi::experimental::ext_oneapi_enable_peer_access(const queue &active, const queue &peer);
+
   /// Constructs a SYCL queue instance using the device returned by an instance
   /// of default_selector.
   ///

sycl/plugins/cuda/pi_cuda.cpp

@@ -5258,6 +5258,19 @@ pi_result cuda_piextUSMEnqueueMemAdvise(pi_queue queue, const void *ptr,
   return result;
 }
 
+pi_result cuda_piextEnablePeer(pi_queue command_queue, pi_queue peer_queue){
+
+ pi_result result = PI_SUCCESS;
+try {
+    ScopedContext active(command_queue->get_context());
+result = PI_CHECK_ERROR(cuCtxEnablePeerAccess(peer_queue->get_context()->get(), 0));
+
+  } catch (pi_result err) {
+    result = err;
+  }
+return result;
+}
+
 /// API to query information about USM allocated pointers
 /// Valid Queries:
 ///   PI_MEM_ALLOC_TYPE returns host/device/shared pi_host_usm value
@@ -5524,6 +5537,7 @@ pi_result piPluginInit(pi_plugin *PluginInit) {
   _PI_CL(piextKernelSetArgSampler, cuda_piextKernelSetArgSampler)
   _PI_CL(piPluginGetLastError, cuda_piPluginGetLastError)
   _PI_CL(piTearDown, cuda_piTearDown)
+  _PI_CL(piextEnablePeer, cuda_piextEnablePeer)
 
 #undef _PI_CL
 

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -8489,6 +8489,13 @@ pi_result piextUSMEnqueueMemAdvise(pi_queue Queue, const void *Ptr,
   return PI_SUCCESS;
 }
 
+pi_result piextEnablePeer(pi_queue command_queue, pi_queue peer_queue){
+
+ pi_result result = PI_SUCCESS;
+
+return result;
+}
+
 /// API to query information about USM allocated pointers.
 /// Valid Queries:
 ///   PI_MEM_ALLOC_TYPE returns host/device/shared pi_usm_type value

sycl/source/queue.cpp

@@ -20,6 +20,22 @@
 namespace sycl {
 __SYCL_INLINE_VER_NAMESPACE(_V1) {
 
+namespace ext {
+namespace oneapi {
+namespace experimental {
+
+bool __SYCL_EXPORT ext_oneapi_enable_peer_access(const queue &active, const queue &peer) {
+  const RT::PiQueue Queue = active.impl->getHandleRef();
+  const detail::plugin &Plugin = active.impl->getPlugin();
+  const RT::PiQueue PeerPi = peer.impl->getHandleRef();
+  Plugin.call<detail::PiApiKind::piextEnablePeer>(Queue, PeerPi );
+  return true;
+}
+
+} // namespace experimental
+} // namespace oneapi
+} // namespace ext
+
 queue::queue(const context &SyclContext, const device_selector &DeviceSelector,
              const async_handler &AsyncHandler, const property_list &PropList) {