1.3.2 Timing accuracy of Fast Models

Fast Models do not model instruction timing accurately. The simulation as a whole has a very accurate concept of timing, but the Code Translation (CT) processors do not claim to dispatch instructions with device-like timing.

In general, a processor issues a set of instructions (a “quantum”) at the same point in simulation time, and then waits for some amount of time before executing the next quantum. The timing is arranged so that the processor averages one instruction per clock tick.

The consequences of this are:

  • The perceived performance of software running on the model differs from real-world software. (In particular, memory accesses and arithmetic operations all take a significant amount of time.)
  • A program might be able to detect the quantized execution behavior of a processor, for example by polling a high-resolution timer.
  • All instructions in a quantum are effectively atomic.

    Note:

    This might mask some race-condition bugs in software.
Non-ConfidentialPDF file icon PDF version100964_1161_00_en
Copyright © 2014–2019 Arm Limited or its affiliates. All rights reserved.