A Guide to the Whetstone CPU Performance Benchmark

This document provides a detailed guide for developers and performance engineers on the openvela system to use the whetstone benchmark tool. whetstone is a classic comprehensive benchmark program designed to accurately evaluate a system's floating-point and integer arithmetic performance by executing a series of standardized computational tasks.

I. Overview

whetstone is a command-line tool used to evaluate a processor's arithmetic performance. It derives a standardized performance score by executing a mixed set of computational tasks, including floating-point operations, integer operations, function calls, and array access.

The test results can help developers:

Quantitatively evaluate the performance of the processor's Floating-Point Unit (FPU).
Analyze the impact of different compiler optimization levels (-O2, -O3, etc.) on code execution efficiency.
Compare arithmetic performance across different hardware platforms or system configurations.

II. System Configuration

Before running the test, you must enable the following Kconfig options in your defconfig file to ensure the whetstone test suite and its dependencies are compiled correctly.

# Compile the Whetstone benchmark tool
CONFIG_BENCHMARK_WHETSTONE=y

# The core of the Whetstone test is floating-point arithmetic, so C library floating-point support must be enabled
CONFIG_LIBC_FLOATINGPOINT=y

III. Usage

The whetstone tool is launched via a simple command-line interface and supports parameters to control the test load.

1. Command Syntax

whetstone [-c <iterations>] [<loops>]

2. Parameter Description

Parameter	Description	Required	Default Value
`[<loops>]`	Module Loop Count. A positional argument that sets the number of execution loops for each internal test module. Increasing this value significantly increases the computational load and execution time of a single test module.	No	1000
`-c <iterations>`	Total Test Rounds. An optional parameter that sets the number of times the entire `whetstone` test suite is repeated.	No	1

3. Execution Examples

Run a standard test: Uses default parameters, with each module looping 1,000 times and the entire test running for 1 round.
```
whetstone
```
Increase the computational load of each module: Each module loops 100,000 times, and the entire test runs for 1 round. This is suitable for scenarios requiring longer run times to obtain a stable average.
```
whetstone 100000
```
Repeat the test multiple times: Each module loops 100,000 times, and the entire test suite is repeated for 10 rounds.
```
whetstone 100000 -c 10
```

IV. Interpreting the Results

After the test is complete, whetstone outputs the test configuration, total time elapsed, and the final performance score.

1. Example Output

ap> whetstone 100000
Loops: 100000, Iterations: 1, Duration: 5 sec.
C Converted Double Precision Whetstones: 2.00 MWIPS

Loops: 100000: Each module executed 100,000 loops.
Iterations: 1: The entire test suite was executed for 1 round.
Duration: 5 sec: Total execution time was 5 seconds.
C Converted Double Precision Whetstones: 2.00 MWIPS: The final performance score is 2.00 MWIPS.

2. Key Metrics Explained

MWIPS / KWIPS
- Meaning: The performance units for whetstone, which stand for MWIPS (Mega Whetstone Instructions Per Second) and KWIPS (Kilo Whetstone Instructions Per Second).
- Calculation: This value is calculated based on a fixed benchmark workload, the number of test loops (loops and -c parameters), and the total execution time. It is a standardized score; the higher the score, the stronger the processor's arithmetic performance.
- Unit Conversion: When the score is below 1 MWIPS (i.e., 1000 KWIPS), the result is displayed in KWIPS.

V. Test Module Details

The whetstone benchmark consists of 11 carefully designed computational modules that comprehensively cover different types of operations:

Modules 1-4: Basic floating-point operations, array operations, and conditional judgments
Module 5: Omitted integer arithmetic module
Module 6: Complex integer arithmetic
Module 7: Trigonometric function calculations (including inverse trigonometric functions)
Module 8: Procedure call test
Module 9: Array indexing test
Module 10: Simple integer arithmetic
Module 11: Chained mathematical function calls

VI. openvela Porting Notes

This version of whetstone has been critically optimized for embedded real-time systems.

Optimized Timer Precision: The original whetstone could only produce results when the test duration reached the second level. The version ported to openvela provides precise output even at the millisecond level. This allows for quick, stable, and accurate performance data even with fewer loop iterations on high-performance embedded CPUs.