[SYCL] use OOO queues and will fall back to use in-order queue if OOO queue creation fails

      1, SYCL runtime will try to create out-of-order queue first.
      2, If it's created successfully, use this out-of-order queue.
      3, If not, fallback to the emulation approach with in-order queue.

Signed-off-by: Chunyang Dai <[email protected]>

Author: cdai2

sycl/include/CL/sycl/detail/queue_impl.hpp

@@ -60,8 +60,9 @@ class queue_impl {
 
   cl_command_queue get() {
     if (m_OpenCLInterop) {
-      CHECK_OCL_CODE(clRetainCommandQueue(m_CommandQueue));
-      return m_CommandQueue;
+      cl_command_queue command_queue = getHandleRef();
+      CHECK_OCL_CODE(clRetainCommandQueue(command_queue));
+      return command_queue;
     }
     throw invalid_object_error(
         "This instance of queue doesn't support OpenCL interoperability");
@@ -120,11 +121,10 @@ class queue_impl {
     m_Exceptions.clear();
   }
 
-  cl_command_queue createQueue() const {
+  cl_command_queue createQueue() {
     cl_command_queue_properties CreationFlags = 0;
 
-    // FPGA RT can't handle out of order queue - create in order queue instead
-    if (!m_Device.is_accelerator()) {
+    if (m_SupportOOO) {
       CreationFlags = CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE;
     }
 
@@ -145,7 +145,15 @@ class queue_impl {
     Queue = clCreateCommandQueue(ClContext, m_Device.get(),
                                           CreationFlags, &Error);
 #endif
-    CHECK_OCL_CODE(Error);
+    // Tf creating out-of-order queue failed and this property is not
+    // supported(for example, on FPGA), it will return
+    // CL_INVALID_QUEUE_PROPERTIES and will try to create in-order queue.
+    if (m_SupportOOO && Error == CL_INVALID_QUEUE_PROPERTIES) {
+      m_SupportOOO = false;
+      Queue = createQueue();
+    } else {
+      CHECK_OCL_CODE(Error);
+    }
     // TODO catch an exception and put it to list of asynchronous exceptions
 
     return Queue;
@@ -171,7 +179,7 @@ class queue_impl {
   }
 
   cl_command_queue &getHandleRef() {
-    if (!m_Device.is_accelerator()) {
+    if (m_SupportOOO) {
       return m_CommandQueue;
     }
 
@@ -218,6 +226,8 @@ class queue_impl {
 
   bool m_OpenCLInterop = false;
   bool m_HostQueue = false;
+  // Assume OOO support by default.
+  bool m_SupportOOO = true;
 };
 
 } // namespace detail

sycl/test/fpga_tests/fpga_queue.cpp

@@ -34,6 +34,18 @@ void GetCLQueue(event sycl_event, std::set<cl_command_queue>& cl_queues) {
   }
 }
 
+int getExpectedQueueNumber(cl_device_id device_id, int default_value) {
+   cl_command_queue_properties reportedProps;
+   cl_int iRet = clGetDeviceInfo(device_id,
+                                 CL_DEVICE_QUEUE_ON_HOST_PROPERTIES,
+                                 sizeof(reportedProps),
+                                 &reportedProps,
+                                 NULL);
+   assert(CL_SUCCESS == iRet && "Failed to obtain queue info from ocl device");
+   return (reportedProps & CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE)
+              ? 1 : default_value;
+}
+
 int main() {
   int data[dataSize] = {0};
 
@@ -98,11 +110,11 @@ int main() {
 
     int result = cl_queues.size();
     device dev = Queue.get_device();
-    int expected_result = dev.is_accelerator() ? 3 : dev.is_host() ? 0 : 1;
+    int expected_result = dev.is_host() ? 0 : getExpectedQueueNumber(dev.get(), 3);
 
     if (expected_result != result) {
       std::cout << "Result Num of queues = " << result << std::endl
-                << "Expected Num of queues = 3" << std::endl;
+                << "Expected Num of queues = "<< expected_result << std::endl;
 
       return -1;
     }
@@ -140,12 +152,11 @@ int main() {
 
     int result = cl_queues.size();
     device dev = Queue.get_device();
-    int expected_result = dev.is_accelerator() ? maxNumQueues :
-                          dev.is_host() ? 0 : 1;
+    int expected_result = dev.is_host() ? 0 : getExpectedQueueNumber(dev.get(), maxNumQueues);
 
     if (expected_result != result) {
       std::cout << "Result Num of queues = " << result << std::endl
-                << "Expected Num of queues = " << maxNumQueues << std::endl;
+                << "Expected Num of queues = " << expected_result << std::endl;
 
       return -1;
     }