Transformer Implementation

This document gives an overview of how the base transformer and computation is implemented, using the NumPy and GPU transformers as examples.

Transformer Creation

The base transformer constructor initializes a set of all computations and results associated with the transformer. As computation objects are created, these sets are populated. Additionally the transformer constructor can build a list of passes to run on the op graph when initialization and transformation is executed.

Specific transformer implementations may use the constructor to initialize code generators (as in the NumPy transformer) or initialize the target device and determine device capabilities (as for the GPU transformer).

Computation Creation

To create a computation, a user calls the transformer’s computation method. This is a relatively lightweight operation that creates a new Computation object and stores it in the set of all computations. The Computation constructor updates the transformer’s set of results and builds a set of all ops that are dependencies of the results by traversing the graph backwards from the result nodes. Computations can only be created through a transformer before the transformer has been initialized. This is partially because the transformer in its current state modifies the graph through passes and tensor description initialization, but this will likely change in the future.

Transformer Initialization

The Transformer.initialize method of the transformer is responsible for running passes to augment the graph, generating code or executables to evaluate ops, allocating buffers and tensors, and initializing tensors. This method can be called manually by the user, but will be automatically called upon the first evaluation of a computation if the user has not manually called it.

Passes and op transformation are called from Transformer._transform_computations. Device buffer and tensor allocation is called from Transformer.allocate_storage which must be implemented by each transformer. Constant tensor initialization is called from Transformer.allocate and other initialization is performed in a special computation called by Transformer.initialize.

Transformer Passes

Transformer passes are run in Transformer._transform_computations here:

def _transform_computations(self):
    Transform computation graphs to a form that can be run.

    # with Op.saved_user_deps():
    # Run passes on the computation graphs

Transformer passes are used to replace ops in the graph, remove ops from the graph, or splice ops into the graph. These passes can be used to simplify the graph (see SimplePrune for an example). Passes can also be used to alter the graph to meet device specific constraints or to optimize ops for exection on specific devices. Currently the only pass that falls into this category is the RequiredTensorShaping pass which reduces the dimensionality of tensors to 2d for reduction elementwise operations and 1d for all other elementwise operations. This pass simplifies the requirements placed on code generation to handle multi-dimensional tensors. In the future, we will likely make this device specific (for example, the GPU kernel generator can handle up to 3 dimensions efficiently).

All passes inherit from the GraphPass class which requires that a child class implement do_pass. Currently all implemented passes are instances of PeepholeGraphPass. A peephole graph pass is a specific type of pass that traverses the graph one node at a time, calling PeepholeGraphPass.visit on each node and builds a mapping of ops to be replaced. Implementors can define visit methods for relevant op types and call PeepholeGraphPass.replace_op to replace the visited op with another op. An example from the SimplePrune pass is shown below:

def visit(self, op):
    x, y = op.args
    rep = None
    if x.is_scalar and x.is_constant:
        if x.const == 0:
            rep = y
    elif y.is_scalar and y.is_constant:
        if y.const == 0:
            rep = x
    if rep is not None:
        self.replace_op(op, rep)

The @visit.on_type(Add) line indicates that this method will be called when an Add op is encountered during graph traversal. This implementation of visit is checking if either of the arguments to Add is 0. Since adding 0 to a value is essentially a no-op, an op meeting this condition can be replaced with its non-zero argument.

Passes present a major opportunity for performance optimization that we plan to exploit in the future. This will likely include device specific fusion of operations that will allow generation of kernels to execute multiple ops at once and buffer sharing that will allow non-overlapping operations to share device memory. These passes would improve execution time and memory usage respectively. We currently have machinery for doing fusion and buffer sharing in ngraph.analysis, but it is not working in the pre-release and has been disabled until we can refactor it to utilize passses.

Intialization Computation

A special initialization computation is created in Transformer._transform_computations which is responsible for executing any initializers attached to graph ops. Initializers are discovered and enumerated in Transformer.ordered_initializers by tranversing the graph and checking for the Op.initializers member. This set of initializers is used to construct the initialization computation

# Collect up all ops from the graph and obtain the init graph
all_ops = OrderedSet(Op.ordered_ops(self.all_results))
init_op = doall(self.ordered_initializers(all_ops))


# create computation which initializes values (called once per session)
self.init_computation = self.computation(init_op, name="init")

This computation is then transformed in the same manner as other computations and run later in Transformer.initialize.

Tensor Description Initialization

Tensor descriptions for all ops are initialized in Transformer.initialize_tensor_descriptions. This calls into the transformer to create DeviceBufferStorage and DeviceTensor instances for each op. Each transformer must define implementations of DeviceBufferStorage and DeviceTensor.

The DeviceBufferStorage class represents a memory allocation on the transformer’s device (for example this will be allocated with PyCUDA for the GPU transformer). This buffer can be used a storage by one or more tensors. When a DeviceBufferStorage object is created, the buffer is not allocated yet, but the object is added to the Transformer.device_buffers member for later allocation.

The DeviceTensor class represents a tensor view on top of a device memory allocation including a base address offset, shape, strides, and data type. A DeviceTensor object is created for every TensorDescription in the graph during Transformer.initialize_tensor_descriptions. When a DeviceTensor object is created, the individual transformer can handle it in multiple ways. The NumPy and GPU transformers both tag DeviceTensor objects to their underlying DeviceBufferStorage objects so that they can be allocated at the same time as the device allocation. Each transformer’s DeviceTensor implementation must support some simple operations including copying to and from NumPy arrays. This is used to set argument values in the graph and get result values from the graph.

After all tensor descriptions are initialized and have created their device buffers and tensors, their allocation is transformed:

for device_buffer in self.device_buffers:

What this means is that the actual allocation of buffers and tensors is transformed into an executable format similar to computations so that it can be run later. This transformed allocation operation is eventually executed by the Transformer.allocate_storage method.

Computation Transformation

Computation objects are finally transformed into an executable format after allocations are transformed in Transformer._transform_computations

for computation in self.computations:

The Computation.transform method first gets the set of all ops needed to evaluation the computation. Since graph passes may have replaced ops by updating their forward pointers, this method will get the fully forwarded set of ops. Then the ops are ordered in such a way that all execution dependencies are met using Digraph.can_reach.

Each transformer implements a Transformer.transform_ordered_ops which accepts a list of ordered ops and transforms them into an executable format. The NumPy transformer implements this by generating a python function containing one or more NumPy calls for each op. Individual ops are handled in the NumPy transformer with the corresponding NumPyCodeGenerator.generate_op implementation. The GPU transformer implements this by generating a GPUKernelGroup containing a set of GPUKernel objects which can be executed to evaluate each op. Individual ops are handled in the GPU transformer with the corresponding GPUKernelGroup.add_kernel implementation or ElementWiseKernel.add_op implementation. The ElementWiseKernel generates CUDA C code to evaluate most op types. Other more complex ops have hand-written GPU kernels such as convolution and GEMM. These are handled in different GPUKernel implementations.

When transformation of computations has finished, the transformer implementation must set the Computation.executor member to either a function or callable object which will serve as the entry point for computation evaluation.

Computation Execution

Computations are executed by calling the Computation.executor member. For the NumPy transformer this is a function pointer to the corresponding function in the generated python NumPy code. For the GPU transformer this is the corresponding GPUKernelGroup object which implements the __call__ method.