env variable control over device selector in DPNPC (#577)

Author: shssf

dpnp/backend/examples/example3.cpp

@@ -42,7 +42,7 @@ int main(int, char**)
 {
     const size_t size = 256;
 
-    dpnp_queue_initialize_c(QueueOptions::CPU_SELECTOR);
+    dpnp_queue_initialize_c();
     std::cout << "SYCL queue is CPU: " << dpnp_queue_is_cpu_c() << std::endl;
 
     int* array1 = (int*)dpnp_memory_alloc_c(size * sizeof(int));

dpnp/backend/include/dpnp_iface.hpp

@@ -68,7 +68,8 @@
 enum class QueueOptions : uint32_t
 {
     CPU_SELECTOR, /**< CPU side execution mode */
-    GPU_SELECTOR  /**< Intel GPU side execution mode */
+    GPU_SELECTOR, /**< Intel GPU side execution mode */
+    AUTO_SELECTOR /**< Automatic selection based on environment variable with @ref CPU_SELECTOR default */
 };
 
 /**
@@ -77,9 +78,9 @@ enum class QueueOptions : uint32_t
  *
  * Global SYCL queue initialization.
  *
- * @param [in]  selector       Select type @ref QueueOptions of the SYCL queue.
+ * @param [in]  selector       Select type @ref QueueOptions of the SYCL queue. Default @ref AUTO_SELECTOR
  */
-INP_DLLEXPORT void dpnp_queue_initialize_c(QueueOptions selector);
+INP_DLLEXPORT void dpnp_queue_initialize_c(QueueOptions selector = QueueOptions::AUTO_SELECTOR);
 
 /**
  * @ingroup BACKEND_API

dpnp/backend/src/queue_sycl.cpp

@@ -35,6 +35,42 @@ cl::sycl::queue* backend_sycl::queue = nullptr;
 #endif
 mkl_rng::mt19937* backend_sycl::rng_engine = nullptr;
 
+static void show_avalable_sycl_devices()
+{
+    const std::vector<cl::sycl::device> devices = cl::sycl::device::get_devices();
+
+    std::cout << "Available SYCL devices:" << std::endl;
+    for (std::vector<cl::sycl::device>::const_iterator it = devices.cbegin(); it != devices.cend(); ++it)
+    {
+        std::cout
+            // not yet implemented error << " " << it->has(sycl::aspect::usm_shared_allocations)  << " "
+            << " - id=" << it->get_info<cl::sycl::info::device::vendor_id>()
+            << ", type=" << static_cast<pi_uint64>(it->get_info<cl::sycl::info::device::device_type>())
+            << ", gws=" << it->get_info<cl::sycl::info::device::max_work_group_size>()
+            << ", cu=" << it->get_info<cl::sycl::info::device::max_compute_units>()
+            << ", name=" << it->get_info<cl::sycl::info::device::name>() << std::endl;
+    }
+}
+
+static cl::sycl::device get_default_sycl_device()
+{
+    int dpnpc_queue_gpu = 0;
+    cl::sycl::device dev = cl::sycl::device(cl::sycl::cpu_selector());
+
+    const char* dpnpc_queue_gpu_var = getenv("DPNPC_QUEUE_GPU");
+    if (dpnpc_queue_gpu_var != NULL)
+    {
+        dpnpc_queue_gpu = atoi(dpnpc_queue_gpu_var);
+    }
+
+    if (dpnpc_queue_gpu)
+    {
+        dev = cl::sycl::device(cl::sycl::gpu_selector());
+    }
+
+    return dev;
+}
+
 /**
  * Function push the SYCL kernels to be linked (final stage of the compilation) for the current queue
  *
@@ -89,14 +125,22 @@ void backend_sycl::backend_sycl_queue_init(QueueOptions selector)
 
     cl::sycl::device dev;
 
+#if not defined(NDEBUG)
+    show_avalable_sycl_devices();
+#endif
+
     if (QueueOptions::CPU_SELECTOR == selector)
     {
         dev = cl::sycl::device(cl::sycl::cpu_selector());
     }
-    else
+    else if (QueueOptions::GPU_SELECTOR == selector)
     {
         dev = cl::sycl::device(cl::sycl::gpu_selector());
     }
+    else
+    {
+        dev = get_default_sycl_device();
+    }
 
     queue = new cl::sycl::queue(dev, exception_handler);