tfp.experimental.auto_batching.stackless.execute

Executes a given program in stackless auto-batching mode.

Compare auto_batching.virtual_machine.execute, which executes the program in full auto-batching mode.

Advantages:

  • Stackless mode is much simpler than full mode.
  • Stackless mode incurs much less overhead, especially in TensorFlow.

Disadvantages:

  • Stackless mode is only compatible with TensorFlow Eager.
  • Stackless mode cannot batch execution across function call boundaries. This is only relevant for recursive functions, and then only if they might converge at different stack depths.

Algorithm:

  • Each function call is executed by its own (recursive) call to _interpret, with a current "active threads" mask.

    • This amounts to borrowing the stack from Python so we don't have to implement it.
    • This is why it is not possible to converge across function call boundaries.
  • The variable environment only has registers (and temporaries and nulls); no expensive full variables, since we are reusing the Python stack.

  • To execute one control flow graph:

    • Maintain (i) a current program counter, (ii) a current active threads mask, and (iii) a priority queue of waiting points, each with a mask giving the logical threads that are waiting to run from there. All these masks should be disjoint.

    • At each time step, execute the basic block indicated by the current program counter, updating only the active threads.

    • Execute FunctionCallOp by recursively invoking _interpret (with only the threads that were active on entry to the FunctionCallOp).

    • If a block ends in a goto, enqueue the target, with the current active thread mask -- all these threads are waiting to resume there.

    • If a block ends in a branch, split the active threads according to the condition value and enqueue two entries, one for the true branch and one for the false branch. This is how divergence happens.

    • At the end of each block (after enqueueing successors), dequeue the smallest program counter, and make active all the threads that were waiting there. This is how re-convergence happens.

    • If the smallest remaining program counter is off the end of the graph, return.

  • Notes: (i) To avoid infinite regress, it's important to avoid actually enqueueing any blocks with an empty waiting thread mask; and (ii) In order to actually take advantage of re-convergence, we take care to coalesce queued entries waiting for the same block (by computing the or of their masks).

This is a reimplementation in TensorFlow Eager of [1].

[1] James Bradbury and Chunli Fu, "Automatic Batching as a Compiler Pass in PyTorch", Workshop on Systems for ML and Open Source Software at NeurIPS 2018.

program A instructions.Program to execute.
backend Object implementing required backend operations.
block_code_cache Dict used to enable caching of defun+XLA across multiple calls to execute. If None is provided, we use a new dict per call to execute which can still achieve caching across depths of the call stack. This caching has no real effect unless calls to backend.wrap_straightline_callable have some effect.
*inputs Input arrays, each of shape [batch_size, e1, ..., eE]. The batch size must be the same for all inputs. The other dimensions must agree with the declared shapes of the variables they will be stored in, but need not in general be the same as one another.

results A list of the output values. Each returned value is an array of shape [batch_size, e1, ..., eE]. The results are returned in the same order as the variables appear in program.out_vars.