matmul + add (fusion)

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

  if (consumer->opcode() == HloOpcode::kDot) {
    // In the general case we call out to optimized "black box" GEMM routines
    // for Dot, which precludes fusion.  However, in very specific cases, we try
    // to fuse Dot operations by generating an elemental dot implementation.
    //
    // We need to be careful and conservative here since any benefit we get from
    // fusion can easily be overshadowed by the overhead of a naive GEMM
    // algorithm in the IR.