Bug 116085 – SunSpider 1.0: Regarding checking result of floating-point number

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Tanaka]

On WiiU platform, RESULTS are not displayed correctly when I validate SunSpider 1.0.
Following each error are displayed and not possible to get total result.

- 3d-morph : ERROR: Invalid test run.
- 3d-raytrace : ERROR: Invalid test run.
- access-nbody : ERROR: Invalid test run.
- math-spectral-norm : ERROR: Invalid test run.

It seems the causes of those errors are checking result value of floating-point number.
In ECMAscript 262, Math.sin and Math.sqrt are not mentioned well.

So it is better to change the contents to allow the margin of error.(example:reduce digit number)

Those problems are confirmed on PSvita and other platform also.

[Filip Pizlo]

(In reply to comment #0)
> On WiiU platform, RESULTS are not displayed correctly when I validate SunSpider 1.0.
> Following each error are displayed and not possible to get total result.
> 
> - 3d-morph : ERROR: Invalid test run.
> - 3d-raytrace : ERROR: Invalid test run.
> - access-nbody : ERROR: Invalid test run.
> - math-spectral-norm : ERROR: Invalid test run.
> 
> It seems the causes of those errors are checking result value of floating-point number.
> In ECMAscript 262, Math.sin and Math.sqrt are not mentioned well.
> 
> So it is better to change the contents to allow the margin of error.(example:reduce digit number)
> 
> Those problems are confirmed on PSvita and other platform also.

Tests that use Math.sin already allow for a large margin of error.

For Math.sqrt, you should provide more details.  While it's true ECMAScript says very little about math precision, I don't think we don't want to go down the path of loosening SunSpider's validation unless we know exactly why we are doing it.

[Tanaka]

(In reply to comment #1)

Thank you for your comment.

> For Math.sqrt, you should provide more details.  While it's true ECMAScript says very little about math precision, I don't think we don't want to go down the path of loosening SunSpider's validation unless we know exactly why we are doing it.

Please check my Math.sqrt investigation as below.

- 3d-raytrace
-- Cause of error: floating-point number error of polynomial including Math.sqrt
-- Test contents : Check string length of Java Script code for canvas
-- expected value: 20970
-- WiiU value    : 21009
-- Proposed      : validate number (not validate string length)

- access-nbody
-- Cause of error: floating-point number error of polynomial including Math.sqrt
-- Test : Total value of calculated result
-- expected value: -1.3524862408537375
-- WiiU value    : -1.3524862408537381 (last number and second last number are not same)
-- Proposed      : validate 10 digit floating point numbers

- math-spectral-norm
-- Cause of error: floating-point number error of polynomial including Math.sqrt
-- Test          : Total value of calculated result
-- expected value: 5.086694231303284
-- WiiU value    : 5.086694231303283 (last number is not same)
-- Proposed      : validate 10 digit floating point numbers

[Tanaka]

Currently, RESULTS are displayed correctly when I validate SunSpider 1.0.2.
So, this proposal should be closed. Thank you.