Replace use of queue mgr (#1735)

* Remove use of deprecated DPCTLQueueMgr_* functions

This is done by reverting to DPNP_LOCAL_QUEUE=1 behavior.

The ultimate goal is to get rid of DPNP_QUEUE throughout
dpnp codebase

* Removed setting DPNP_LOCAL_QUEUE in tests

Since DPNP_LOCAL_QUEUE has been removed.

* Removed uses of device_context

* Removed trailing whitespace

* Replaced tab with spaces

* Use backend_sycl_singleton, remove backend_sycl

* Removed pre-preprocessor fenced out dead code

Renamed backend_sycl_singleton back to backend_sycl

Removed dpnp_queue_initialize_c function and
QueueOptions enum as no longer used. Removed uses of these
from examples and code base

* Formatting suggested by pre-commit hooks

* Addressed issues raised by @diptorupd

---------

Co-authored-by: Anton <[email protected]>

Author: oleksandr-pavlyk

doc/dpctl.rst

@@ -10,20 +10,23 @@ Interplay with the Data Parallel Control Library
 An example below demonstrates how the Data Parallel Extension for NumPy* can be
 easily combined with the device management interface provided by dpctl package.
 
-Literally, the SYCL* queue manager interface from the dpctl package allows
-to set an input queue as the currently usable queue inside the context
-manager's scope. This way an array creation function from the dpnp package
-which is defined inside the context will allocate the data using that queue.
-
 .. code-block:: python
   :linenos:
 
-  import dpctl
-  import dpnp as np
+      import dpctl
+      import dpnp
+
+      d = dpctl.select_cpu_device()
+      x = dpnp.array([1, 2, 3], device=d)
+      s = dpnp.sum(x)
+
+      y = dpnp.linspace(0, dpnp.pi, num=10**6, device="gpu")
+      f = 1 + y * dpnp.sin(y)
+
+      # locate argument where function attains global maximum
+      max_arg = x[dpnp.argmax(f)]
+      max_val = dpnp.max(f)
 
-  with dpctl.device_context("opencl:gpu"):
-      x = np.array([1, 2, 3])
-      s = np.sum(x)
 
 For more information please refer to `Data Parallel Control Library`_
 documentation.

dpnp/backend/examples/example10.cpp

@@ -50,8 +50,6 @@ void test_dpnp_random_normal(const size_t size,
     double dev_time_used = 0.0;
     double sum_dev_time_used = 0.0;
 
-    dpnp_queue_initialize_c(QueueOptions::GPU_SELECTOR);
-
     double *result = (double *)dpnp_memory_alloc_c(size * sizeof(double));
 
     dpnp_rng_srand_c(seed); // TODO: will move

dpnp/backend/examples/example3.cpp

@@ -44,7 +44,6 @@ int main(int, char **)
 {
     const size_t size = 256;
 
-    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/examples/example5.cpp

@@ -52,8 +52,6 @@ int main(int, char **)
 {
     const size_t size = 256;
 
-    dpnp_queue_initialize_c(QueueOptions::CPU_SELECTOR);
-
     double *result = (double *)dpnp_memory_alloc_c(size * sizeof(double));
 
     size_t seed = 10;

dpnp/backend/examples/example7.cpp

@@ -45,8 +45,6 @@ int main(int, char **)
     const size_t size = 2;
     size_t len = size * size;
 
-    dpnp_queue_initialize_c(QueueOptions::CPU_SELECTOR);
-
     float *array = (float *)dpnp_memory_alloc_c(len * sizeof(float));
     float *result1 = (float *)dpnp_memory_alloc_c(size * sizeof(float));
     float *result2 = (float *)dpnp_memory_alloc_c(len * sizeof(float));

dpnp/backend/examples/example8.cpp

@@ -42,8 +42,6 @@ int main(int, char **)
 {
     const size_t size = 16;
 
-    dpnp_queue_initialize_c(QueueOptions::GPU_SELECTOR);
-
     double *array = (double *)dpnp_memory_alloc_c(size * sizeof(double));
     long *result = (long *)dpnp_memory_alloc_c(size * sizeof(long));
 

dpnp/backend/examples/example9.cpp

@@ -46,8 +46,6 @@ int main(int, char **)
     long result = 0;
     long result_verification = 0;
 
-    dpnp_queue_initialize_c(QueueOptions::CPU_SELECTOR);
-
     long *array =
         reinterpret_cast<long *>(dpnp_memory_alloc_c(size * sizeof(long)));
 

dpnp/backend/examples/example_bs.cpp

@@ -229,7 +229,6 @@ int main(int, char **)
     const double RISK_FREE = 0.1;
     const double VOLATILITY = 0.2;
 
-    dpnp_queue_initialize_c(QueueOptions::GPU_SELECTOR);
     std::cout << "SYCL queue is CPU: " << dpnp_queue_is_cpu_c() << std::endl;
 
     double *price = (double *)dpnp_memory_alloc_c(SIZE * sizeof(double));

dpnp/backend/include/dpnp_iface.hpp

@@ -68,33 +68,6 @@ typedef ssize_t shape_elem_type;
  * @}
  */
 
-/**
- * @ingroup BACKEND_API
- * @brief SYCL queue initialization selector.
- *
- * The structure defines the parameters that are used for the library
- * initialization by @ref dpnp_queue_initialize_c "dpnp_queue_initialize".
- */
-enum class QueueOptions : uint32_t
-{
-    CPU_SELECTOR, /**< CPU side execution mode */
-    GPU_SELECTOR, /**< Intel GPU side execution mode */
-    AUTO_SELECTOR /**< Automatic selection based on environment variable with
-                     @ref CPU_SELECTOR default */
-};
-
-/**
- * @ingroup BACKEND_API
- * @brief SYCL queue initialization.
- *
- * Global SYCL queue initialization.
- *
- * @param [in]  selector       Select type @ref QueueOptions of the SYCL queue.
- * Default @ref AUTO_SELECTOR
- */
-INP_DLLEXPORT void dpnp_queue_initialize_c(
-    QueueOptions selector = QueueOptions::AUTO_SELECTOR);
-
 /**
  * @ingroup BACKEND_API
  * @brief SYCL queue device status.
@@ -112,8 +85,7 @@ INP_DLLEXPORT size_t dpnp_queue_is_cpu_c();
  * @param [in]  size_in_bytes  Number of bytes for requested memory allocation.
  * @param [in]  q_ref          Reference to SYCL queue.
  *
- * @return  A pointer to newly created memory on @ref dpnp_queue_initialize_c
- * "initialized SYCL device".
+ * @return  A pointer to newly created memory on SYCL device.
  */
 INP_DLLEXPORT char *dpnp_memory_alloc_c(DPCTLSyclQueueRef q_ref,
                                         size_t size_in_bytes);

dpnp/backend/kernels/dpnp_krnl_random.cpp

@@ -70,7 +70,8 @@ static VSLStreamStatePtr get_rng_stream()
 
 void dpnp_rng_srand_c(size_t seed)
 {
-    backend_sycl::backend_sycl_rng_engine_init(seed);
+    auto &be = backend_sycl::get();
+    be.set_rng_engines_seed(seed);
     set_rng_stream(seed);
 }
 

dpnp/backend/src/queue_sycl.cpp

@@ -31,13 +31,7 @@
 #include "dpnp_utils.hpp"
 #include "queue_sycl.hpp"
 
-#if defined(DPNP_LOCAL_QUEUE)
-sycl::queue *backend_sycl::queue = nullptr;
-#endif
-mkl_rng::mt19937 *backend_sycl::rng_engine = nullptr;
-mkl_rng::mcg59 *backend_sycl::rng_mcg59_engine = nullptr;
-
-static void dpnpc_show_mathlib_version()
+[[maybe_unused]] static void dpnpc_show_mathlib_version()
 {
 #if 1
     const int len = 256;
@@ -61,8 +55,8 @@ static void dpnpc_show_mathlib_version()
 #endif
 }
 
-#if (not defined(NDEBUG)) && defined(DPNP_LOCAL_QUEUE)
-static void show_available_sycl_devices()
+#if (not defined(NDEBUG))
+[[maybe_unused]] static void show_available_sycl_devices()
 {
     const std::vector<sycl::device> devices = sycl::device::get_devices();
 
@@ -86,25 +80,6 @@ static void show_available_sycl_devices()
 }
 #endif
 
-#if defined(DPNP_LOCAL_QUEUE)
-static sycl::device get_default_sycl_device()
-{
-    int dpnpc_queue_gpu = 0;
-    sycl::device dev = sycl::device(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 = sycl::device(sycl::gpu_selector());
-    }
-
-    return dev;
-}
-#endif
-
 #if defined(DPNPC_TOUCH_KERNEL_TO_LINK)
 /**
  * Function push the SYCL kernels to be linked (final stage of the compilation)
@@ -135,117 +110,8 @@ static long dpnp_kernels_link()
 }
 #endif
 
-#if defined(DPNP_LOCAL_QUEUE)
-// Catch asynchronous exceptions
-static void exception_handler(sycl::exception_list exceptions)
-{
-    for (std::exception_ptr const &e : exceptions) {
-        try {
-            std::rethrow_exception(e);
-        } catch (sycl::exception const &e) {
-            std::cout << "DPNP. Caught asynchronous SYCL exception:\n"
-                      << e.what() << std::endl;
-        }
-    }
-};
-#endif
-
-void backend_sycl::backend_sycl_queue_init(QueueOptions selector)
-{
-#if defined(DPNP_LOCAL_QUEUE)
-    std::chrono::high_resolution_clock::time_point t1 =
-        std::chrono::high_resolution_clock::now();
-
-    if (queue) {
-        backend_sycl::destroy();
-    }
-
-    sycl::device dev;
-
-#if not defined(NDEBUG)
-    show_available_sycl_devices();
-#endif
-
-    if (QueueOptions::CPU_SELECTOR == selector) {
-        dev = sycl::device(sycl::cpu_selector());
-    }
-    else if (QueueOptions::GPU_SELECTOR == selector) {
-        dev = sycl::device(sycl::gpu_selector());
-    }
-    else {
-        dev = get_default_sycl_device();
-    }
-
-    if (is_verbose_mode()) {
-        sycl::property_list properties{
-            sycl::property::queue::enable_profiling()};
-        queue = new sycl::queue(dev, exception_handler, properties);
-    }
-    else {
-        queue = new sycl::queue(dev, exception_handler);
-    }
-
-    std::chrono::high_resolution_clock::time_point t2 =
-        std::chrono::high_resolution_clock::now();
-    std::chrono::duration<double> time_queue_init =
-        std::chrono::duration_cast<std::chrono::duration<double>>(t2 - t1);
-#else
-    (void)selector;
-#endif
-
-    std::chrono::high_resolution_clock::time_point t3 =
-        std::chrono::high_resolution_clock::now();
-#if defined(DPNPC_TOUCH_KERNEL_TO_LINK)
-    // Remove pre-link kernel library at startup time
-    dpnp_kernels_link();
-#endif
-    std::chrono::high_resolution_clock::time_point t4 =
-        std::chrono::high_resolution_clock::now();
-    std::chrono::duration<double> time_kernels_link =
-        std::chrono::duration_cast<std::chrono::duration<double>>(t4 - t3);
-
-    std::cout << "Running on: "
-              << DPNP_QUEUE.get_device().get_info<sycl::info::device::name>()
-              << "\n";
-#if defined(DPNP_LOCAL_QUEUE)
-    std::cout << "queue initialization time: " << time_queue_init.count()
-              << " (sec.)\n";
-#else
-    std::cout << "DPCtrl SYCL queue used\n";
-#endif
-    std::cout << "SYCL kernels link time: " << time_kernels_link.count()
-              << " (sec.)\n";
-    dpnpc_show_mathlib_version();
-
-    std::cout << std::endl;
-}
-
-bool backend_sycl::backend_sycl_is_cpu()
-{
-    sycl::queue &qptr = get_queue();
-
-    if (qptr.get_device().is_cpu()) {
-        return true;
-    }
-
-    return false;
-}
-
-void backend_sycl::backend_sycl_rng_engine_init(size_t seed)
-{
-    if (rng_engine) {
-        backend_sycl::destroy_rng_engine();
-    }
-    rng_engine = new mkl_rng::mt19937(DPNP_QUEUE, seed);
-    rng_mcg59_engine = new mkl_rng::mcg59(DPNP_QUEUE, seed);
-}
-
-void dpnp_queue_initialize_c(QueueOptions selector)
-{
-    backend_sycl::backend_sycl_queue_init(selector);
-}
-
 size_t dpnp_queue_is_cpu_c()
 {
-    return backend_sycl::backend_sycl_is_cpu();
+    const auto &be = backend_sycl::get();
+    return be.backend_sycl_is_cpu();
 }