AutoCounter: Profiling with Out-of-Band Performance Counter Collection

FireSim can provide visibility into a simulated CPU’s architectural and microarchitectural state over the course of execution through the use of counters. These are similar to performance counters provided by processor vendors, and more general counters provided by architectural simulators.

This functionality is provided by the AutoCounter feature (introduced in our FirePerf paper at ASPLOS 2020), and can be used for profiling and debugging. Since AutoCounter injects counters only in simulation (unlike target-level performance counters), these counters do not affect the behavior of the simulated machine, no matter how often they are sampled.

Chisel Interface

AutoCounter enables the addition of ad-hoc counters using the PerfCounter object in the midas.targetutils package. PerfCounters counters can be added in one of two modes:

  1. Accumulate, using the standard PerfCounter.apply method. Here the annotated UInt (1 or more bits) is added to a 64b accumulation register: the target is treated as representing an N-bit UInt and will increment the counter by a value between [0, 2^n - 1] per cycle.

  2. Identity, using the PerfCounter.identity method. Here the annotated UInt is sampled directly. This can be used to annotate a sample with values are not accumulator-like (e.g., a PC), and permits the user to define more complex instrumentation logic in the target itself.

We give examples of using PerfCounter below:

// A standard boolean event. Increments by 1 or 0 every local clock cycle.
midas.targetutils.PerfCounter(en_clock, "gate_clock", "Core clock gated")

// A multibit example. If the core can retire three isntructions per cycle,
// encode this as a two-bit unit. Extra-width is OK but the encoding to the UInt
// (e.g., doing a pop count), must be done by the user.
midas.targetutils.PerfCounter(insns_ret, "iret", "Instructions retired")

// An identity value. Note: the pc here must be <= 64b wide.
midas.targetutils.PerfCounter.identity(pc, "pc", "The value of the program counter at the time of a sample")

See the PerfCounter Scala API docs for more detail about the Chisel-side interface.

Enabling AutoCounter in Golden Gate

By default, annotated events are not synthesized into AutoCounters. To enable AutoCounter when compiling a design, prepend the WithAutoCounter config to your PLATFORM_CONFIG. During compilation, Golden Gate will print the signals it is generating counters for.

Rocket Chip Cover Functions

The cover function is applied to various signals in the Rocket Chip generator repository to mark points of interest (i.e., interesting signals) in the RTL. Tools are free to provide their own implementation of this function to process these signals as they wish. In FireSim, these functions can be used as a hook for automatic generation of counters.

Since cover functions are embedded throughout the code of Rocket Chip (and possibly other code repositories), AutoCounter provides a filtering mechanism based on module granularity. As such, only cover functions that appear within selected modules will generate counters.

The filtered modules can be indicated using one of two methods:

  1. An annotation attached to the module for which cover functions should be turned into AutoCounters. The annotation requires a ModuleTarget which can be pointed to any module in the design. Alternatively, the current module can be annotated as follows:

class SomeModule(implicit p: Parameters) extends Module
{
  chisel3.experimental.annotate(AutoCounterCoverModuleAnnotation(
      Module.currentModule.get.toTarget))
}
  1. An input file with a list of module names. This input file is named autocounter-covermodules.txt, and includes a list of module names separated by new lines (no commas).

AutoCounter Runtime Parameters

AutoCounter currently takes a single runtime configurable parameter, defined under the autocounter: section in the config_runtime.yaml file. The read_rate parameter defines the rate at which the counters should be read, and is measured in target-cycles of the base target-clock (clock 0 produced by the ClockBridge). Hence, if the read_rate is defined to be 100 and the tile frequency is 2x the base clock (ex., which may drive the uncore), the simulator will read and print the values of the counters every 200 core-clock cycles. If the core-domain clock is the base clock, it would do so every 100 cycles. By default, the read_rate is set to 0 cycles, which disables AutoCounter.

autocounter:
    # read counters every 100 cycles
    read_rate: 100

Note

AutoCounter is designed as a coarse-grained observability mechanism, as sampling each counter requires two (blocking) MMIO reads (each read takes O(100) ns on EC2 F1). As a result sampling at intervals less than O(10000) cycles may adversely affect simulation performance for large numbers of counters. If you intend on reading counters at a finer granularity, consider using synthesizable printfs.

AutoCounter CSV Output Format

AutoCounter output files are CSVs generated in the working directory where the simulator was invoked (this applies to metasimulators too), with the default names AUTOCOUNTERFILE<i>.csv, one per clock domain. The CSV output format is depicted below, assuming a sampling period of N base clock cycles.

AutoCounter CSV Format

version

version number

Clock Domain Name

name

Base Multiplier

M

Base Divisor

N

label

local_cycle

label0

label1

labelN

description

local clock cycle

desc0

desc1

descN

type

Accumulate

type0

type1

typeN

event width

1

width0

width1

widthN

accumulator width

64

64

64

64

N

cycle @ time N

value0 @ tN

value1 @ tN

value @ tN

kN

cycle @ time kN

value0 @ tkN

value1 @ tkN

valueN @ tkN

Column Notes:

  1. Each column beyond the first two corresponds to a PerfCounter instance in the clock domain.

  2. Column 0 past the header corresponds to the base clock cycle of the sample.

  3. The local_cycle counter (column 1) is implemented as an always enabled single-bit event, and increments even when the target is under reset.

Row Notes:

  1. Header row 0: autocounter csv format version, an integer.

  2. Header row 1: clock domain information.

  3. Header row 2: the label parameter provided to PerfCounter suffixed with the instance path.

  4. Header row 3: the description parameter provided to PerfCounter. Quoted.

  5. Header row 4: the width of the field annotated in the target.

  6. Header row 5: the width of the accumulation register. Not configurable, but makes it clear when to expect rollover.

  7. Header row 6: indicates the accumulation scheme. Can be “Identity” or “Accumulate”.

  8. Sample row 0: sampled values at the bitwidth of the accumulation register.

  9. Sample row k: ditto above, k * N base cycles later

Using TracerV Trigger with AutoCounter

In order to collect AutoCounter results from only from a particular region of interest in the simulation, AutoCounter has been integrated with TracerV triggers. See the Setting a TracerV Trigger section for more information.

AutoCounter using Synthesizable Printfs

The AutoCounter transformation in Golden Gate includes an event-driven mode that uses Synthesizable Printfs (see Printf Synthesis: Capturing RTL printf Calls when Running on the FPGA) to export counter results as they are updated rather than sampling them periodically with a dedicated Bridge. This mode can be enabled by prepending the WithAutoCounterCoverPrintf config to your PLATFORM_CONFIG instead of WithAutoCounterCover. Based on the selected event mode the printfs will have the following runtime behavior:

  • Accumulate: On a non-zero increment, the local cycle count and the new counter value are printed. This produces a series of prints with monotonically increasingly values.

  • Identity: On a transition of the annotated target, the local cycle count and the new value are printed. Thus a target that transitions every cycle will produce printf traffic every cycle.

This mode may be useful for temporally fine-grained observation of counters. The counter values will be printed to the same output stream as other synthesizable printfs. This mode uses considerably more FPGA resources per counter, and may consume considerable amounts of DMA bandwidth (since it prints every cycle a counter increments), which may adversly affect simulation performance (increased FMR).

Reset & Timing Considerations

  • Events and identity values provided while under local reset, or while the GlobalResetCondition asserted, are zero-ed out. Similarly, printfs that might otherwise be active under a reset are masked out.

  • The sampling period in slower clock domains is currently calculated using a truncating division of the period in the base clock domain. Thus, when the base clock period can not be cleanly divided, samples in the slower clock domain will gradually fall out of phase with samples in the base clock domain. In all cases, the “local_cycle” column is most accurate measure of sample time.