[WIP]

Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>

dali/c_api_2/data_objects.h

@@ -60,6 +60,22 @@ class TensorListInterface : public RefCountedObject {
   virtual cudaEvent_t GetOrCreateReadyEvent() = 0;
 };
 
+struct TensorListDeleter {
+  daliDeleter_t deleter;
+  AccessOrder deletion_order;
+
+  void operator()(void *data) {
+    if (deleter.delete_buffer) {
+      cudaStream_t stream = deletion_order.stream();
+      deleter.delete_buffer(deleter.deleter_ctx, data,
+                            deletion_order.is_device() ? &stream : nullptr);
+    }
+    if (deleter.destroy_context) {
+      deleter.destroy_context(deleter.destroy_context);
+    }
+  }
+};
+
 template <typename Backend>
 class TensorListWrapper : public TensorListInterface {
  public:
@@ -91,10 +107,7 @@ class TensorListWrapper : public TensorListInterface {
     if (!deleter.delete_buffer && !deleter.destroy_context) {
       buffer.reset(buffer, [](void *){});
     } else {
-      buffer.reset(buffer, [deleter](void *p) {
-        if (deleter.delete_buffer)
-          deleter.delete_buffer(deleter.deleter_ctx, p, nullptr);
-      });
+      buffer.reset(buffer, TensorListDeleter{deleter, order()});
     }
     for (int i = 0; i < num_samples; i++) {
       TensorShape<> sample_shape(make_cspan(&shapes[i*ndim]. ndim));
@@ -105,17 +118,17 @@ class TensorListWrapper : public TensorListInterface {
         sample_data = static_cast<char *>(data) + next_offset;
         next_offset += volme(sample_shape) * element_size;
       }
-
     }
-
   }
 
   virtual void AttachSamples(
       int num_samples,
       int ndim,
       daliDataType_t dtype,
       const daliTensorDesc_t *samples,
-      const daliDeleter_t *sample_deleters) = 0;
+      const daliDeleter_t *sample_deleters) {
+
+  }
 
   virtual daliBufferPlacement_t GetBufferPlacement() const = 0;