tageBuilder

YEAR: 2025 |

Python-based sim/eval tool for TAGE-like branch predictors.

Banner

Overview

tageBuilder is a Python-based simulator and performance evaluation tool for TAGE-like branch predictors. It offers configurable parameters for modeling TAGE predictors and is optimized for performance with Numba and NumPy. Predictor performance can be evaluated with 200+ branch trace datasets from the 2016 Championship Branch Prediction (CBP2016) environment and generates detailed performance metrics.

TAGE predictors are state-of-the-art branch predictor design. If you are curious about their details, check out resources.

Features

Customizable Predictor Model:
Configure aspects such as tagged components (number, history length, tag width) and base predictor settings (prediction bits, hysteresis bits) via a YAML spec file.

The tool does not model non-tagged components (except for the base predictor) such as the loop predictor and the statistical corrector.

A predictor with 12 tagged components with total space of 212Kb scored accuracy of 95%> and MPKI of 1~3 for most traces.
Optimized Simulation:
Uses Numba JIT and NumPy vectorized operations to process ~700,000 branch sequences per second per simulation process.
Batch Processing:
Supports batch prediction and updates for efficient data flow.
Real-World Trace Integration:
Works with BT9-formatted branch trace files from CBP2016, reverse-engineered to meet the tool’s requirements.
Comprehensive Evaluations:
Generates metrics and visualizations including MPKI, accuracy graphs, memory usage charts, misprediction statistics, branch class analysis, and CSV outputs for further analysis.
Parallel Processing:
Runs multiple traces in parallel processes to speed up evaluations.

Technical Details

currently adding more details

Configuration

tageBuilder uses a YAML configuration file to define predictor parameters. This allows users to easily adjust the following:

Number and configuration of tagged components
Base predictor settings (e.g., prediction bits, hysteresis bits)
Global predictor parameters (e.g., path history length, usefulness bit duration)

Below is an example predictor spec with one base bimodal predictor and 4 tagged predictors.

global_config:
    phist_len: 16
    u_duration_log: 19
tables:
    - name: base
      isBase: true
      hist_len: 0
      num_pred_entries_log: 14
      num_hyst_entries_log: 12
      tag_width: 0

    - name: t1
      isBase: false
      hist_len: 8
      num_pred_entries_log: 10
      num_hyst_entries_log: 0
      tag_width: 8

    - name: t2
      isBase: false
      hist_len: 12
      num_pred_entries_log: 10
      num_hyst_entries_log: 0
      tag_width: 9

    - name: t3
      isBase: false
      hist_len: 18
      num_pred_entries_log: 10
      num_hyst_entries_log: 0
      tag_width: 9

    - name: t4
      isBase: false
      hist_len: 27
      num_pred_entries_log: 10
      num_hyst_entries_log: 0
      tag_width: 10