0001224: GlossyReflection kernel can compute a NaN Vector

View Issue Details [ Jump to Notes ]

[ Issue History ] [ Print ]

Project

Category

View Status

Date Submitted

Last Update

0001224

LuxRender

Core

public

2012-04-23 11:14

2012-09-11 06:11

Reporter

raun

Assigned To

Lord Crc

Priority

normal

Severity

minor

Reproducibility

sometimes

Status

closed

Resolution

fixed

Platform

Windows - OpenCL

Win 7 64

OS Version

Product Version

0.8

Target Version

Fixed in Version

1.0RC1

Summary

0001224: GlossyReflection kernel can compute a NaN Vector

Description

In the interesting case, the two inputs to pow() are base = 0x3f7ff9df = slightly less than 1.0f and exponent = 0x39d0a3c4 = a pretty small positive value.

- Windows calculator computes this as a value close to and slightly less than 1.0f, or 0x3F7FFFFF.
- A conformant version of pow() can compute this as a value slightly above 1.0f, or 0x3f800001

const float cosTheta = pow(1.f - u1, exponent); // cosTheta = slightly >1.0f
const float sinTheta = sqrt(1.f - cosTheta * cosTheta); // sinTheta = NaN!
const float x = cos(phi) * sinTheta; // x = NaN!
const float y = sin(phi) * sinTheta; // y = NaN!
const float z = cosTheta;

Eventually, our entire ray direction is nothing but NaNs. This causes the entire QBVH to get walked as part of Intersect(), which means >200K iterations vs. 0000045:0000100 if the ray direction is non-NaN. That obviously impacts performance.

If we clamp the results of pow() to [0.0f, 1.0f]:

const float cosTheta = min( 1.f, pow(1.f - u1, exponent) );

everything is fine.

Tags

No tags attached.

Mercurial Changeset #

6033098d1c43

Requires Documentation Update

Requires Exporter Update

Attached Files

Relationships

Notes
(0003381) SATtva (developer) 2012-04-23 20:39	Can you recheck with 1.0RC1 or the current trunk?

(0003382) raun (reporter) 2012-04-25 09:42	Appears patched in this changeset: author Asbjørn Heid Wed Apr 25 09:05:49 2012 +0200 (9 hours ago) changeset 1048 6033098d1c43 parent 1046 6f1d6b9c48bd child 1049 271407fbc7b9 471 void GlossyReflection(const Vector wo, Vector wi, const float exponent, 472 const Vector shadeN, const float u0, const float u1) { 473 const float phi = 2.f M_PI * u0; 474 const float cosTheta = pow(1.f - u1, exponent); 475 const float sinTheta = sqrt(max(0.f, 1.f - cosTheta * cosTheta)); 476 const float x = cos(phi) * sinTheta; 477 const float y = sin(phi) * sinTheta; 478 const float z = cosTheta; This fixes the NaN result from sqrt, but its still possible that costTheta is > 1.0f. I'm not sure if this is a problem since the vector is probably normalized eventually. Thanks for the quick response!

Issue History
Date Modified	Username	Field	Change
2012-04-23 11:14	raun	New Issue
2012-04-23 20:39	SATtva	Note Added: 0003381
2012-04-25 09:42	raun	Note Added: 0003382
2012-05-12 13:57	Lord Crc	Mercurial Changeset #	=> 6033098d1c43
2012-05-12 13:57	Lord Crc	Status	new => resolved
2012-05-12 13:57	Lord Crc	Fixed in Version	=> 1.0RC1
2012-05-12 13:57	Lord Crc	Resolution	open => fixed
2012-05-12 13:57	Lord Crc	Assigned To	=> Lord Crc
2012-09-11 06:11	jeanphi	Status	resolved => closed