Don't emit calls to Eigen for the dot operator. Emit an LLVM loop instead to provide visibility to Polly.

Author: arpith-jacob

tensorflow/compiler/xla/service/cpu/dot_op_emitter.cc

@@ -1013,6 +1013,7 @@ bool PotentiallyImplementedAsEigenDot(const HloInstruction& hlo) {
 DotInLlvmIrProfitable ProfitableToImplementDotInUntiledLlvmIr(
  const HloInstruction& dot) {
  if (dot.opcode() == HloOpcode::kDot && dot.shape().dimensions_size() == 2) {
+#if 0
  const Shape& result_shape = dot.shape();
  // kReductionDimensionThresholdBytes was chosen to be 1/4 of a typical L1
  // cache line size, so that we can have the reduction dimension of both the
@@ -1048,6 +1049,8 @@ DotInLlvmIrProfitable ProfitableToImplementDotInUntiledLlvmIr(
  ? DotInLlvmIrProfitable::kWithColumnMajorRhs
  : DotInLlvmIrProfitable::kYes;
  }
+#endif
+ return DotInLlvmIrProfitable::kYes;
  }
  return DotInLlvmIrProfitable::kNo;
 }