[SYCL][CUDA] Add support for ALLOC_HOST_PTR (#1997)

Signed-off-by: Ruyman Reyes <[email protected]>

Author: Ruyk

sycl/include/CL/sycl/detail/pi.h

@@ -508,6 +508,7 @@ constexpr pi_mem_flags PI_MEM_FLAGS_ACCESS_RW = CL_MEM_READ_WRITE;
 // Host pointer
 constexpr pi_mem_flags PI_MEM_FLAGS_HOST_PTR_USE = CL_MEM_USE_HOST_PTR;
 constexpr pi_mem_flags PI_MEM_FLAGS_HOST_PTR_COPY = CL_MEM_COPY_HOST_PTR;
+constexpr pi_mem_flags PI_MEM_FLAGS_HOST_PTR_ALLOC = CL_MEM_ALLOC_HOST_PTR;
 
 // NOTE: queue properties are implemented this way to better support bit
 // manipulations

sycl/plugins/cuda/pi_cuda.cpp

@@ -1650,6 +1650,10 @@ 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::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::alloc_mode::alloc_host_ptr;
     } else {
       retErr = PI_CHECK_ERROR(cuMemAlloc(&ptr, size));
       if (flags & PI_MEM_FLAGS_HOST_PTR_COPY) {
@@ -1721,6 +1725,8 @@ pi_result cuda_piMemRelease(pi_mem memObj) {
       case _pi_mem::alloc_mode::use_host_ptr:
         ret = PI_CHECK_ERROR(cuMemHostUnregister(uniqueMemObj->hostPtr_));
         break;
+      case _pi_mem::alloc_mode::alloc_host_ptr:
+        ret = PI_CHECK_ERROR(cuMemFreeHost(uniqueMemObj->hostPtr_));
       };
     }
 

sycl/plugins/cuda/pi_cuda.hpp

@@ -197,8 +197,14 @@ struct _pi_mem {
    * 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 } allocMode_;
+  enum class alloc_mode {
+    classic,
+    use_host_ptr,
+    copy_in,
+    alloc_host_ptr
+  } allocMode_;
 
   _pi_mem(pi_context ctxt, pi_mem parent, alloc_mode mode, CUdeviceptr ptr, void *host_ptr,
           size_t size)

sycl/unittests/pi/cuda/test_mem_obj.cpp

@@ -73,6 +73,18 @@ TEST_F(CudaTestMemObj, piMemBufferCreateSimple) {
             PI_SUCCESS);
 }
 
+TEST_F(CudaTestMemObj, piMemBufferAllocHost) {
+  const size_t memSize = 1024u;
+  pi_mem memObj;
+  ASSERT_EQ((plugin.call_nocheck<detail::PiApiKind::piMemBufferCreate>(
+                context_, PI_MEM_FLAGS_ACCESS_RW | PI_MEM_FLAGS_HOST_PTR_ALLOC,
+                memSize, nullptr, &memObj)),
+            PI_SUCCESS);
+
+  ASSERT_EQ((plugin.call_nocheck<detail::PiApiKind::piMemRelease>(memObj)),
+            PI_SUCCESS);
+}
+
 TEST_F(CudaTestMemObj, piMemBufferCreateNoActiveContext) {
   const size_t memSize = 1024u;
   // Context has been destroyed