CUDA provider split

Author: KFilipek

src/provider/provider_cuda.c

@@ -374,20 +374,6 @@ static umf_result_t cu_memory_provider_initialize(const void *params,
     snprintf(cu_provider->name, sizeof(cu_provider->name), "%s",
              cu_params->name);
 
-    // CUDA alloc functions doesn't allow to provide user alignment - get the
-    // minimum one from the driver
-    size_t min_alignment = 0;
-    CUmemAllocationProp allocProps = {0};
-    allocProps.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-    allocProps.type = CU_MEM_ALLOCATION_TYPE_PINNED;
-    allocProps.location.id = cu_params->cuda_device_handle;
-    CUresult cu_result = g_cu_ops.cuMemGetAllocationGranularity(
-        &min_alignment, &allocProps, CU_MEM_ALLOC_GRANULARITY_MINIMUM);
-    if (cu_result != CUDA_SUCCESS) {
-        umf_ba_global_free(cu_provider);
-        return cu2umf_result(cu_result);
-    }
-
     cu_provider->context = cu_params->cuda_context_handle;
     cu_provider->device = cu_params->cuda_device_handle;
     cu_provider->memory_type = cu_params->memory_type;
@@ -396,14 +382,28 @@ static umf_result_t cu_memory_provider_initialize(const void *params,
     // If the memory type is shared (CUDA managed), the allocation flags must
     // be set. NOTE: we do not check here if the flags are valid -
     // this will be done by CUDA runtime.
-    if (cu_params->memory_type == UMF_MEMORY_TYPE_SHARED &&
+    if (cu_provider->memory_type == UMF_MEMORY_TYPE_SHARED &&
         cu_params->alloc_flags == 0) {
         // the default setting is CU_MEM_ATTACH_GLOBAL
         cu_provider->alloc_flags = CU_MEM_ATTACH_GLOBAL;
     } else {
         cu_provider->alloc_flags = cu_params->alloc_flags;
     }
 
+    // CUDA alloc functions doesn't allow to provide user alignment - get the
+    // minimum one from the driver
+    size_t min_alignment = 0;
+    CUmemAllocationProp allocProps = {0};
+    allocProps.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+    allocProps.type = CU_MEM_ALLOCATION_TYPE_PINNED;
+    allocProps.location.id = cu_params->cuda_device_handle;
+    CUresult cu_result = g_cu_ops.cuMemGetAllocationGranularity(
+        &min_alignment, &allocProps, CU_MEM_ALLOC_GRANULARITY_MINIMUM);
+    if (cu_result != CUDA_SUCCESS) {
+        umf_ba_global_free(cu_provider);
+        return cu2umf_result(cu_result);
+    }
+
     *provider = cu_provider;
 
     return UMF_RESULT_SUCCESS;