Re-introduce changes from host alloc pull request (intel#1997)

Author: nyalloc

sycl/plugins/cuda/pi_cuda.cpp

@@ -1651,8 +1651,15 @@ pi_result cuda_piMemBufferCreate(pi_context context, pi_mem_flags flags,
           cuMemHostRegister(host_ptr, size, CU_MEMHOSTREGISTER_DEVICEMAP));
       retErr = PI_CHECK_ERROR(cuMemHostGetDevicePointer(&ptr, host_ptr, 0));
       allocMode = _pi_mem::mem_::buffer_mem_::alloc_mode::use_host_ptr;
+    } else if (flags & PI_MEM_FLAGS_HOST_PTR_ALLOC) {
+      retErr = PI_CHECK_ERROR(cuMemAllocHost(&host_ptr, size));
+      retErr = PI_CHECK_ERROR(cuMemHostGetDevicePointer(&ptr, host_ptr, 0));
+      allocMode = _pi_mem::mem_::buffer_mem_::alloc_mode::alloc_host_ptr;
     } else {
       retErr = PI_CHECK_ERROR(cuMemAlloc(&ptr, size));
+      if (flags & PI_MEM_FLAGS_HOST_PTR_COPY) {
+        allocMode = _pi_mem::mem_::buffer_mem_::alloc_mode::copy_in;
+      }
     }
 
     if (retErr == PI_SUCCESS) {
@@ -1715,13 +1722,16 @@ pi_result cuda_piMemRelease(pi_mem memObj) {
 
     if (memObj->mem_type_ == _pi_mem::mem_type::buffer) {
       switch (uniqueMemObj->mem_.buffer_mem_.allocMode_) {
+      case _pi_mem::mem_::buffer_mem_::alloc_mode::copy_in:
       case _pi_mem::mem_::buffer_mem_::alloc_mode::classic:
         ret = PI_CHECK_ERROR(cuMemFree(uniqueMemObj->mem_.buffer_mem_.ptr_));
         break;
       case _pi_mem::mem_::buffer_mem_::alloc_mode::use_host_ptr:
         ret = PI_CHECK_ERROR(
             cuMemHostUnregister(uniqueMemObj->mem_.buffer_mem_.hostPtr_));
         break;
+      case _pi_mem::mem_::buffer_mem_::alloc_mode::alloc_host_ptr:
+        ret = PI_CHECK_ERROR(cuMemFreeHost(uniqueMemObj->mem_.buffer_mem_.hostPtr_));
       };
     } else if (memObj->mem_type_ == _pi_mem::mem_type::surface) {
       ret = PI_CHECK_ERROR(

sycl/plugins/cuda/pi_cuda.hpp

@@ -199,7 +199,20 @@ struct _pi_mem {
       size_t mapOffset_;
       void *mapPtr_;
       cl_map_flags mapFlags_;
-      enum class alloc_mode { classic, use_host_ptr } allocMode_;
+
+      /** alloc_mode
+       * classic: Just a normal buffer allocated on the device via cuda malloc
+       * use_host_ptr: Use an address on the host for the device
+       * copy_in: The data for the device comes from the host but the host
+       pointer is not available later for re-use
+       * alloc_host_ptr: Uses pinned-memory allocation
+      */
+      enum class alloc_mode {
+        classic,
+        use_host_ptr,
+        copy_in,
+        alloc_host_ptr
+      } allocMode_;
 
       native_type get() const noexcept { return ptr_; }
 
@@ -237,6 +250,7 @@ struct _pi_mem {
         return mapFlags_;
       }
     } buffer_mem_;
+    
     struct surface_mem_ {
       CUarray array_;
       CUsurfObject surfObj_;