Cache generated LLVM IR for GEBP
authorSanjoy Das <sanjoy@google.com>
Thu, 24 May 2018 00:49:42 +0000 (17:49 -0700)
committerTensorFlower Gardener <gardener@tensorflow.org>
Thu, 24 May 2018 00:52:23 +0000 (17:52 -0700)
commit335d52c17644f417bc53abe4ef87ead9de01ad6d
tree94a11f269aa54fe37f1523a72049d1d18af92c2a
parent0f5f23e876be89dc2a389be078289d3028ae6503
Cache generated LLVM IR for GEBP

After this change all generated GEBPs with the same shape will share a single
llvm::Function.

This is NFC for any actual workloads because the GEBP emitter isn't exercised by
normal code-paths yet.

PiperOrigin-RevId: 197820606
tensorflow/compiler/xla/service/cpu/dot_op_emitter.cc