Feature/memory zero copy

backends/include/dppl_sycl_device_interface.h

@@ -203,4 +203,15 @@ DPPLDevice_GetVendorName (__dppl_keep const DPPLSyclDeviceRef DRef);
 DPPL_API
 bool DPPLDevice_IsHostUnifiedMemory (__dppl_keep const DPPLSyclDeviceRef DRef);
 
+/*!
+ * @brief Checks if two DPPLSyclDeviceRef objects point to the same
+ * sycl::device.
+ *
+ * @param    DevRef1       First opaque pointer to the sycl device.
+ * @param    DevRef2       Second opaque pointer to the sycl device.
+ * @return   True if the underlying sycl::device are same, false otherwise.
+ */
+DPPL_API
+bool DPPLDevice_AreEq (__dppl_keep const DPPLSyclDeviceRef DevRef1,
+                       __dppl_keep const DPPLSyclDeviceRef DevRef2);
 DPPL_C_EXTERN_C_END

backends/include/dppl_sycl_queue_interface.h

@@ -199,4 +199,30 @@ DPPL_API
 void DPPLQueue_Memcpy (__dppl_keep const DPPLSyclQueueRef QRef,
                        void *Dest, const void *Src, size_t Count);
 
+/*!
+ * @brief C-API wrapper for sycl::queue::prefetch, the function waits on an event
+ * till the prefetch operation completes.
+ *
+ * @param    QRef           An opaque pointer to the sycl queue.
+ * @param    Ptr            An USM pointer to memory.
+ * @param    Count          A number of bytes to prefetch.
+ */
+DPPL_API
+void DPPLQueue_Prefetch (__dppl_keep DPPLSyclQueueRef QRef,
+			 const void *Ptr, size_t Count);
+
+/*!
+ * @brief C-API wrapper for sycl::queue::mem_advise, the function waits on an event
+ * till the operation completes.
+ *
+ * @param    QRef           An opaque pointer to the sycl queue.
+ * @param    Ptr            An USM pointer to memory.
+ * @param    Count          A number of bytes to prefetch.
+ * @param    Advice         Device-defined advice for the specified allocation. 
+ *                           A value of 0 reverts the advice for Ptr to the default behavior.
+ */
+DPPL_API
+void DPPLQueue_MemAdvise (__dppl_keep DPPLSyclQueueRef QRef,
+			  const void *Ptr, size_t Count, int Advice);
+
 DPPL_C_EXTERN_C_END

backends/include/dppl_sycl_queue_manager.h

@@ -158,4 +158,20 @@ DPPLQueueMgr_PushQueue (DPPLSyclBackendType BETy,
 DPPL_API
 void DPPLQueueMgr_PopQueue ();
 
+
+/*!
+ * @brief Creates a new instance of SYCL queue from SYCL context and
+ * SYCL device.
+ *
+ * The instance is not placed into queue manager. The user assumes
+ * ownership of the queue reference and should deallocate it using
+ * DPPLQueue_Delete.
+ *
+ */
+DPPL_API
+__dppl_give DPPLSyclQueueRef
+DPPLQueueMgr_GetQueueFromContextAndDevice(__dppl_keep DPPLSyclContextRef CRef,
+					  __dppl_keep DPPLSyclDeviceRef DRef);
+
+
 DPPL_C_EXTERN_C_END

backends/include/dppl_sycl_usm_interface.h

@@ -42,6 +42,16 @@ DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_shared (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
+/*!
+ * @brief Crete USM shared memory.
+ *
+ * @return The pointer to USM shared memory with requested alignment.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_shared (size_t alignment, size_t size,
+			  __dppl_keep const DPPLSyclQueueRef QRef);
+
 /*!
  * @brief Crete USM host memory.
  *
@@ -51,6 +61,16 @@ DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_host (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
+/*!
+ * @brief Crete USM host memory.
+ *
+ * @return The pointer to USM host memory with requested alignment.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_host (size_t alignment, size_t size,
+			__dppl_keep const DPPLSyclQueueRef QRef);
+
 /*!
  * @brief Crete USM device memory.
  *
@@ -60,6 +80,16 @@ DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_device (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
+/*!
+ * @brief Crete USM device memory.
+ *
+ * @return The pointer to USM device memory with requested alignment.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_device (size_t alignment, size_t size,
+			  __dppl_keep const DPPLSyclQueueRef QRef);
+
 /*!
  * @brief Free USM memory.
  *
@@ -86,4 +116,13 @@ const char *
 DPPLUSM_GetPointerType (__dppl_keep const DPPLSyclUSMRef MRef,
                         __dppl_keep const DPPLSyclContextRef CRef);
 
+/*!
+ * @brief Get the device associated with USM pointer.
+ *
+ * @return A DPPLSyclDeviceRef pointer to the sycl device.
+ */
+DPPL_API
+DPPLSyclDeviceRef
+DPPLUSM_GetPointerDevice (__dppl_keep const DPPLSyclUSMRef MRef,
+			  __dppl_keep const DPPLSyclContextRef CRef);
 DPPL_C_EXTERN_C_END

backends/source/dppl_sycl_device_interface.cpp

@@ -261,3 +261,12 @@ bool DPPLDevice_IsHostUnifiedMemory (__dppl_keep const DPPLSyclDeviceRef DRef)
     }
     return false;
 }
+
+bool DPPLDevice_AreEq(__dppl_keep const DPPLSyclDeviceRef DevRef1,
+		      __dppl_keep const DPPLSyclDeviceRef DevRef2)
+{
+    if(!(DevRef1 && DevRef2))
+        // \todo handle error
+        return false;
+    return (*unwrap(DevRef1) == *unwrap(DevRef2));
+}

backends/source/dppl_sycl_queue_interface.cpp

@@ -290,10 +290,28 @@ DPPLQueue_Wait (__dppl_keep DPPLSyclQueueRef QRef)
     SyclQueue->wait();
 }
 
-void DPPLQueue_Memcpy (__dppl_take const DPPLSyclQueueRef QRef,
+void DPPLQueue_Memcpy (__dppl_keep const DPPLSyclQueueRef QRef,
                        void *Dest, const void *Src, size_t Count)
 {
     auto Q = unwrap(QRef);
     auto event = Q->memcpy(Dest, Src, Count);
     event.wait();
 }
+
+void
+DPPLQueue_Prefetch (__dppl_keep DPPLSyclQueueRef QRef,
+		    const void *Ptr, size_t Count)
+{
+    auto Q = unwrap(QRef);
+    auto event = Q->prefetch(Ptr, Count);
+    event.wait();
+}
+
+void
+DPPLQueue_MemAdvise (__dppl_keep DPPLSyclQueueRef QRef,
+		     const void *Ptr, size_t Count, int Advice)
+{
+    auto Q = unwrap(QRef);
+    auto event = Q->mem_advise(Ptr, Count, static_cast<pi_mem_advice>(Advice));
+    event.wait();
+}

backends/source/dppl_sycl_queue_manager.cpp

@@ -40,6 +40,8 @@ namespace
 
 // Create wrappers for C Binding types (see CBindingWrapping.h).
 DEFINE_SIMPLE_CONVERSION_FUNCTIONS(queue, DPPLSyclQueueRef)
+DEFINE_SIMPLE_CONVERSION_FUNCTIONS(device, DPPLSyclDeviceRef)
+DEFINE_SIMPLE_CONVERSION_FUNCTIONS(context, DPPLSyclContextRef)
 
 /*!
  * @brief A helper class to support the DPPLSyclQueuemanager.
@@ -534,3 +536,18 @@ void DPPLQueueMgr_PopQueue ()
 {
     QMgrHelper::popSyclQueue();
 }
+
+/*!
+ * The function constructs a new SYCL queue instance from SYCL conext and
+ * SYCL device.
+ */
+DPPLSyclQueueRef
+DPPLQueueMgr_GetQueueFromContextAndDevice(__dppl_keep DPPLSyclContextRef CRef,
+					  __dppl_keep DPPLSyclDeviceRef DRef)
+{
+    auto dev = unwrap(DRef);
+    auto ctx = unwrap(CRef);
+    auto q = queue(*ctx, *dev);
+
+    return wrap(new queue(q));
+}

backends/source/dppl_sycl_usm_interface.cpp

@@ -25,6 +25,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "dppl_sycl_usm_interface.h"
+#include "dppl_sycl_device_interface.h"
 #include "Support/CBindingWrapping.h"
 
 #include <CL/sycl.hpp>                /* SYCL headers   */
@@ -35,6 +36,7 @@ namespace
 {
 // Create wrappers for C Binding types (see CBindingWrapping.h).
 DEFINE_SIMPLE_CONVERSION_FUNCTIONS(queue, DPPLSyclQueueRef)
+DEFINE_SIMPLE_CONVERSION_FUNCTIONS(device, DPPLSyclDeviceRef)
 DEFINE_SIMPLE_CONVERSION_FUNCTIONS(context, DPPLSyclContextRef)
 DEFINE_SIMPLE_CONVERSION_FUNCTIONS(void, DPPLSyclUSMRef)
 
@@ -48,6 +50,15 @@ DPPLmalloc_shared (size_t size, __dppl_keep const DPPLSyclQueueRef QRef)
     return wrap(Ptr);
 }
 
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_shared (size_t alignment, size_t size,
+			  __dppl_keep const DPPLSyclQueueRef QRef)
+{
+    auto Q = unwrap(QRef);
+    auto Ptr = aligned_alloc_shared(alignment, size, *Q);
+    return wrap(Ptr);
+}
+
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_host (size_t size, __dppl_keep const DPPLSyclQueueRef QRef)
 {
@@ -56,6 +67,15 @@ DPPLmalloc_host (size_t size, __dppl_keep const DPPLSyclQueueRef QRef)
     return wrap(Ptr);
 }
 
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_host (size_t alignment, size_t size,
+			__dppl_keep const DPPLSyclQueueRef QRef)
+{
+    auto Q = unwrap(QRef);
+    auto Ptr = aligned_alloc_host(alignment, size, *Q);
+    return wrap(Ptr);
+}
+
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_device (size_t size, __dppl_keep const DPPLSyclQueueRef QRef)
 {
@@ -64,6 +84,15 @@ DPPLmalloc_device (size_t size, __dppl_keep const DPPLSyclQueueRef QRef)
     return wrap(Ptr);
 }
 
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_device (size_t alignment, size_t size,
+			  __dppl_keep const DPPLSyclQueueRef QRef)
+{
+    auto Q = unwrap(QRef);
+    auto Ptr = aligned_alloc_device(alignment, size, *Q);
+    return wrap(Ptr);
+}
+
 void DPPLfree_with_queue (__dppl_take DPPLSyclUSMRef MRef,
                           __dppl_keep const DPPLSyclQueueRef QRef)
 {
@@ -99,3 +128,15 @@ DPPLUSM_GetPointerType (__dppl_keep const DPPLSyclUSMRef MRef,
             return "unknown";
     }
 }
+
+DPPLSyclDeviceRef
+DPPLUSM_GetPointerDevice(__dppl_keep const DPPLSyclUSMRef MRef,
+			 __dppl_keep const DPPLSyclContextRef CRef)
+{
+    auto Ptr = unwrap(MRef);
+    auto C = unwrap(CRef);
+
+    auto Dev = get_pointer_device(Ptr, *C);
+
+    return wrap(new device(Dev));
+}

backends/tests/CMakeLists.txt

@@ -29,6 +29,7 @@ else()
     test_sycl_program_interface
     test_sycl_queue_interface
     test_sycl_queue_manager
+    test_sycl_usm_interface
   )
 
   # Copy the spir-v input files to test build directory

backends/tests/test_sycl_queue_manager.cpp

@@ -244,6 +244,22 @@ TEST_F (TestDPPLSyclQueueManager, CheckIsCurrentQueue2)
     DPPLQueueMgr_PopQueue();
 }
 
+TEST_F (TestDPPLSyclQueueManager, CreateQueueFromDeviceAndContext)
+{
+    auto Q = DPPLQueueMgr_GetCurrentQueue();
+    auto D = DPPLQueue_GetDevice(Q);
+    auto C = DPPLQueue_GetContext(Q);
+
+    auto Q2 = DPPLQueueMgr_GetQueueFromContextAndDevice(C, D);
+    auto D2 = DPPLQueue_GetDevice(Q2);
+    auto C2 = DPPLQueue_GetContext(Q2);
+
+    EXPECT_TRUE(DPPLDevice_AreEq(D, D2));
+    EXPECT_TRUE(DPPLContext_AreEq(C, C2));
+
+    DPPLQueue_Delete(Q2);
+}
+
 int
 main (int argc, char** argv)
 {