[GitHub] [incubator-mxnet] ptrendx commented on a change in pull request #19209: [BUGFIX]Fix cuDNN dropout reproducibility

[GitBox <gi...@apache.org>]

ptrendx commented on a change in pull request #19209:
URL: https://github.com/apache/incubator-mxnet/pull/19209#discussion_r492986716



##########
File path: src/resource.cc
##########
@@ -428,37 +486,28 @@ void* Resource::get_host_space_internal(size_t size) const {
 void Resource::get_cudnn_dropout_desc(
     cudnnDropoutDescriptor_t *dropout_desc,
     mshadow::Stream<gpu> *stream,
-    const float dropout, uint64_t seed,
+    const float dropout,
     const std::string &name) const {
 
   CHECK_EQ(req.type, ResourceRequest::kCuDNNDropoutDesc);
   auto state_space = static_cast<resource::SpaceAllocator*>(ptr_);
   CHECK_EQ(state_space->ctx.dev_id, stream->dev_id)
-    << "The device id of cuDNN dropout state space doesn't match that from stream.";
-  CUDNN_CALL(cudnnCreateDropoutDescriptor(dropout_desc));
+    << "The device id of cudnn dropout state space doesn't match that from stream.";
   if (dropout <= 0) {
     CUDNN_CALL(cudnnSetDropoutDescriptor(*dropout_desc, stream->dnn_handle_,
                                          dropout,
                                          nullptr,
                                          0, seed));
-  } else if (!state_space->handle.size) {
-    // not initialized yet.
-    size_t dropout_state_size;
-    CUDNN_CALL(cudnnDropoutGetStatesSize(stream->dnn_handle_, &dropout_state_size));
-    // reserve GPU space
-    Storage::Get()->DirectFree(Storage::Get()->Alloc(dropout_state_size, state_space->ctx));
-    CUDNN_CALL(cudnnSetDropoutDescriptor(*dropout_desc, stream->dnn_handle_,
-                                         dropout,
-                                         state_space->GetSpace(dropout_state_size, name),
-                                         dropout_state_size, seed));
   } else {
+    CHECK(state_space->handle.size > 0)
+      << "CUDNN dropout descriptor was not initialized yet!";
     // cudnnRestoreDropoutDescriptor() introduced with cuDNN v7
     STATIC_ASSERT_CUDNN_VERSION_GE(7000);
     CUDNN_CALL(cudnnRestoreDropoutDescriptor(*dropout_desc, stream->dnn_handle_,
                                              dropout,
                                              state_space->handle.dptr,
                                              state_space->handle.size,
-                                             seed));
+                                             0));

Review comment:
       It does not matter. From cuDNN API reference:
   ```
   seed
   Input. Seed used in prior call to cudnnSetDropoutDescriptor() that initialized states buffer. Using a different seed from this has no effect. A change of seed, and subsequent update to random number generator states can be achieved by calling cudnnSetDropoutDescriptor().
   ```




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org