CODEWITHSUNDEEP
tensorflow
Minjie Wang
Minjie Wang

Reputation: 1

How XLA express matrix_solve

I browsed through XLA's document and cannot find a good way to express more complicated matrix operations like matrix_solve, matrix_triangular_solve, cholesky and so on. How XLA handles this? I know there is a catch-call operation "CallCustom" but just wonder about better ways.

Upvotes: 0

Views: 147

Answers (1)

Todd Wang
Todd Wang

Reputation: 66

In general, the intention is for the actual computation to be specified in regular TensorFlow. Then you turn on XLA either via Just-In-Time compilation (https://www.tensorflow.org/performance/xla/jit), or Ahead-Of-Time compilation (.../xla/tfcompile).

In terms of underlying support for matrix solvers, note that in addition to typical dense matrix operations, XLA does support some control flow primitives. See https://www.tensorflow.org/performance/xla/operation_semantics paying attention to the while loop construct (#while), and how to select output from different choices (#select).

I haven't worked out whether this will yield a great result, but at a high-level it seems like the fundamental pieces are there.

(Sorry for the abbreviated links; I can't seem to post more than 2)

Upvotes: 1

Related Questions