Merge pull request #152 from IntelPython/feature/memory-zero-copy

Feature/memory zero copy

Author: oleksandr-pavlyk

CHANGELOG.md

@@ -4,6 +4,7 @@ All notable changes to this project will be documented in this file.
 ## [Unreleased]
 ### Added
 - Device descriptors "max_compute_units", "max_work_item_dimensions", "max_work_item_sizes", "max_work_group_size", "max_num_sub_groups" and "aspects" for int64 atomics inside dpctl C API and inside the dpctl.SyclDevice class.
+- MemoryUSM* classes moved to `dpctl.memory` module, added support for aligned allocation, added support for `prefetch` and `mem_advise` (sychronous) methods, implemented `copy_to_host`, `copy_from_host` and `copy_from_device` methods, pickling support, and zero-copy interoperability with Python objects which implement `__sycl_usm_array_inerface__` protocol.
 
 ### Removed
 - The Legacy OpenCL interface.

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 +158,25 @@ 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.
+ *
+ * @param    CRef           Sycl context reference
+ * @param    DRef           Sycl device reference
+ *
+ * @return A copy of the sycl::queue created from given context and device
+ * references.
+ */
+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

@@ -34,35 +34,94 @@
 DPPL_C_EXTERN_C_BEGIN
 
 /*!
- * @brief Crete USM shared memory.
+ * @brief Create USM shared memory.
  *
- * @return The pointer to USM shared memory.
+ * @param    size     Number of bytes to allocate
+ * @param    QRef     Sycl queue reference to use in allocation
+ *
+ * @return The pointer to USM shared memory. On failure, returns nullptr.
  */
 DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_shared (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
 /*!
- * @brief Crete USM host memory.
+ * @brief Create USM shared memory.
+ *
+ * @param  alignment   Allocation's byte alignment
+ * @param  size        Number of bytes to allocate
+ * @param  QRef        Sycl queue reference to use in allocation
  *
- * @return The pointer to USM host memory.
+ * @return The pointer to USM shared memory with the requested alignment.
+ * On failure, returns nullptr.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_shared (size_t alignment, size_t size,
+                          __dppl_keep const DPPLSyclQueueRef QRef);
+
+/*!
+ * @brief Create USM host memory.
+ *
+ * @param    size     Number of bytes to allocate
+ * @param    QRef     Sycl queue reference to use in allocation
+ *
+ * @return The pointer to USM host memory. On failure, returns nullptr.
  */
 DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_host (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
 /*!
- * @brief Crete USM device memory.
+ * @brief Create USM host memory.
+ *
+ * @param  alignment   Allocation's byte alignment
+ * @param  size        Number of bytes to allocate
+ * @param  QRef        Sycl queue reference to use in allocation
  *
- * @return The pointer to USM device memory.
+ * @return The pointer to USM host memory with the requested alignment.
+ * On failure, returns nullptr.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_host (size_t alignment, size_t size,
+                        __dppl_keep const DPPLSyclQueueRef QRef);
+
+/*!
+ * @brief Create USM device memory.
+ *
+ * @param    size     Number of bytes to allocate
+ * @param    QRef     Sycl queue reference to use in allocation
+ *
+ * @return The pointer to USM device memory. On failure, returns nullptr.
  */
 DPPL_API
 __dppl_give DPPLSyclUSMRef
 DPPLmalloc_device (size_t size, __dppl_keep const DPPLSyclQueueRef QRef);
 
+/*!
+ * @brief Create USM device memory.
+ *
+ * @param    alignment   Allocation's byte alignment
+ * @param    size        Number of bytes to allocate
+ * @param    QRef        Sycl queue reference to use in allocation
+ *
+ * @return The pointer to USM device memory with requested alignment.
+ * On failure, returns nullptr.
+ */
+DPPL_API
+__dppl_give DPPLSyclUSMRef
+DPPLaligned_alloc_device (size_t alignment, size_t size,
+                          __dppl_keep const DPPLSyclQueueRef QRef);
+
 /*!
  * @brief Free USM memory.
  *
+ * @param   MRef      USM pointer to free
+ * @param   QRef      Sycl queue reference to use.
+ *
+ * USM pointer must have been allocated using the same context as the one
+ * used to construct the queue.
  */
 DPPL_API
 void DPPLfree_with_queue (__dppl_take DPPLSyclUSMRef MRef,
@@ -79,11 +138,26 @@ void DPPLfree_with_context (__dppl_take DPPLSyclUSMRef MRef,
 /*!
  * @brief Get pointer type.
  *
+ * @param    MRef      USM Memory 
+ * @param    CRef      Sycl context reference associated with the pointer
+ *
  * @return "host", "device", "shared" or "unknown"
  */
 DPPL_API
 const char *
 DPPLUSM_GetPointerType (__dppl_keep const DPPLSyclUSMRef MRef,
                         __dppl_keep const DPPLSyclContextRef CRef);
 
+/*!
+ * @brief Get the device associated with USM pointer.
+ *
+ * @param  MRef    USM pointer
+ * @param  CRef    Sycl context reference associated with the 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_platform_interface.cpp

@@ -41,7 +41,7 @@ get_set_of_non_hostbackends ()
 {
     std::set<DPPLSyclBackendType> be_set;
     for (auto p : platform::get_platforms()) {
-		if(p.is_host())
+        if(p.is_host())
             continue;
         auto be = p.get_backend();
         switch (be)
@@ -155,12 +155,12 @@ void DPPLPlatform_DumpInfo ()
 */
 size_t DPPLPlatform_GetNumNonHostPlatforms ()
 {
-	auto nNonHostPlatforms = 0ul;
-	for (auto &p : platform::get_platforms()) {
-		if (p.is_host())
-			continue;
-		++nNonHostPlatforms;
-	}
+    auto nNonHostPlatforms = 0ul;
+    for (auto &p : platform::get_platforms()) {
+        if (p.is_host())
+            continue;
+        ++nNonHostPlatforms;
+    }
     return nNonHostPlatforms;
 }
 

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.
@@ -99,7 +101,7 @@ class QMgrHelper
     {
         QVec *active_queues;
         try {
-            auto def_device = std::move(default_selector().select_device());
+            auto def_device { default_selector().select_device() };
             auto BE = def_device.get_platform().get_backend();
             auto DevTy = def_device.get_info<info::device::device_type>();
 
@@ -534,3 +536,17 @@ 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);
+
+    return wrap(new queue(*ctx, *dev));
+}

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));
+}