MXNet can integrate with many different kinds of backend libraries, including TVM, MKLDNN, TensorRT, Intel nGraph and more. These backend in general support a limited number of operators, and thus running computation in a model usually involves in interaction between accelerator backend-supported operators and MXNet operators.
These accelerators backend libraries share some common requirements:
- TVM , MKLDNN and nGraph uses customized data formats. Interaction between these accelerators backends with MXNet requires data format conversion.
- TVM, MKLDNN, TensorRT and nGraph fuses operators.
Integration with these accelerators backends should happen in the granularity of subgraphs instead of in the granularity of operators. To fuse operators, it's obvious that we need to divide a graph into subgraphs so that the operators in a subgraph can be fused into a single operator. To handle customized data formats, we should partition a computation graph into subgraphs as well. Each subgraph contains only TVM, MKLDNN or ngraph operators. In this way, MXNet converts data formats only when entering such a subgraph and the operators inside a subgraph handle format conversion themselves if necessary. This makes interaction of TVM and MKLDNN with MXNet much easier. Neither the MXNet executor nor the MXNet operators need to deal with customized data formats.
Even though invoking these libraries from MXNet requires similar steps, the partitioning rule and the subgraph execution of these accelerators backends can be different. As such, we define the following interface for accelerators backends to customize graph partitioning and subgraph execution inside an operator.
...
Step 1: graph partition
Graph partitioning is to traverse a computation graph and group operators into subgraphs based on certain rules. There already exists an TVM fuse pass in NNVM, which groups operators into subgraphs based on certain rules (e.g., convolution followed by element-wise operations). This graph partitioner is TVM-specific. It doesn't work for other acceleratorsbackends. For example, MKLDNN operator fusion requires a partitioner that finds subgraphs with specific patterns (e.g., convolution followed by batchnorm, followed by activation, etc).
Regardless of the diverse partitioning requirements, we assume all graph partitioning shares the following requirements:
...
We provide a default selector called ContainOpSelector, which extracts a subgraph with operators supported by an acceleratora backend. This selector or its variant is commonly needed by most of the acceleratorsbackends.
When a subgraph is found, the partitioning algorithm invokes SubgraphProperty::CreateSubgraphNode to create a new node for the subgraph and connects the new node back to the original graph to replace the subgraph. The subgraph passed to CreateSubgraphNode contains shape/dtype/storage information of the input nodes. This is important because acceleratorsbackends, such as TVM and TensorRT, need these information to compile and select the best kernel. CreateSubgraphNode allows any customization on the subgraph and the node, which gives users many opportunities to optimize the subgraph. For example, some acceleratorsbackends, such as TVM and MKLDNN, can perform another level of graph partitioning in CreateSubgraphNode for operator fusion.
To perform graph partitioning, we attach a graph property (a class that implements SubgraphProperty) and invoke PartitionGraph.
...
Step 2: subgraph operator (function call)
Although there are two levels of graph partitioning, we only need to handle one level of subgraphs in the executor because the subgraphs in the second level are fused into operators. We can execute these subgraphs inside special operators, which is specific to the acceleratorbackend.
- TVM execution operator: loads a subgraph from a TVM compiled binary, a graph JSON file and weight arrays, and executes the subgraph composed of fused operators. We can first use the TVM executor to execute the subgraph, but in the future we should use the MXNet executor because MXNet executes operators in multiple threads, which is useful for task parallelism. The operator needs to convert all output NDArrays of the subgraph to the default format.
- MKLDNN execution operator: gets a subgraph from the first step and runs operators in the MXNet executors. Like TVM operator, this operator also needs to convert all output NDArrays of the subgraph to the default format.
- TensorRT has its engine for executing the optimized subgraph.
- nGraph execution operators: it's similar to TensorRT.
To customize the subgraph execution, an accelerator a backend needs to provide their own operator implementation and attach the operator to the subgraph node when SubgraphProperty::CreateSubgraphNode is called. The subgraph operator should be a stateful operator and contain a computation graph. We provide a default subgraph operator implementation (“_subgraph_op”) that executes operators with MXNet Executor.
For fast inference in TVM and MKLDNN, the subgraph operators need to maintain a copy of weight arrays (similar to the closure of a function). In this way, we can convert the data format of the weight arrays and cache the array inside the subgraph operator to avoid any redundant format conversion. The original weight arrays will still be part of the inputs of the subgraph operator. Even though the weight arrays are normally not modified, we still need to handle this case correctly. One solution is to maintain a version number for the var of an NDArray, which is increased by one whenever the NDArray is modified in the execution engine. We can use the version number to determine the weight arrays have been modified whenever the subgraph operator is invoked.
The benefit of invoking a subgraph inside an operator
Introducing a subgraph operator for TVM and MKLDNN may sound like unnecessary complexity. It actually significantly reduces the complexity of the integration. By using the subgraph operator, we can completely isolate TVM operators and MKLDNN operators from MXNet operators as well as the default MXNet memory planning. Inside the subgraph operators, we don't need to deal with data format conversion and can use a completely different memory plan for the subgraph.
...