tag:blogger.com,1999:blog-8350257063773144600.post4274516644562201385..comments2024-03-29T04:22:20.467-07:00Comments on Pete Shirley's Graphics Blog: Making your BVH fasterPeter Shirleyhttp://www.blogger.com/profile/17871569418798062417noreply@blogger.comBlogger8125tag:blogger.com,1999:blog-8350257063773144600.post-36233081395428704032019-12-10T08:59:28.541-08:002019-12-10T08:59:28.541-08:00Hey Peter,
Looks like the image you uploaded for ...Hey Peter,<br /><br />Looks like the image you uploaded for Martin Lambers' example is dead. Any chance of updating it? Cheers!Matthttps://www.blogger.com/profile/05927581152681614380noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-73478359913841042302019-10-23T09:07:39.326-07:002019-10-23T09:07:39.326-07:00This is the necessary and interesting information ...This is the necessary and interesting information for my work. You wrote, accessible and understandable.<a href="https://custom-paper-writing.com/blog/essay-assignment/" rel="nofollow">https://custom-paper-writing.com/blog/essay-assignment/</a> Correct tips, thanks for sharing this information.paulsmith198914@gmail.comhttps://www.blogger.com/profile/01097862152097877690noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-89969587913489473222019-04-28T00:40:14.305-07:002019-04-28T00:40:14.305-07:00Great post..It is very informative and effective h...Great post..It is very informative and effective here...Thanks for sharing with us...<br /><a href="https://clippingexpertasia.com/image-manipulation" rel="nofollow">image manipulation service</a>Image Manipulation Expert https://www.blogger.com/profile/03140252658474128213noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-18828361554141711392019-03-18T03:48:40.203-07:002019-03-18T03:48:40.203-07:00Extending your SAH point: I wrote a blog post abou...Extending your SAH point: I wrote a blog post about it at https//link.medium.com/gwvS7KvG9URomanWichehttps://www.blogger.com/profile/06320967596486379411noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-37979721000468382432019-03-13T02:14:55.838-07:002019-03-13T02:14:55.838-07:00Good advice on NVIDIA: make sure that nodes and pr...Good advice on NVIDIA: make sure that nodes and prims are loaded from memory using 128bit LD instructions. Proper alignment and data layout is key to this. Can mean an extra 10% saved.<br /><br />Also: recursion hurts here, but that’s a different topic.Stefan Zellmannhttps://www.blogger.com/profile/01631222976811388837noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-27421237272521435682019-03-12T11:45:30.213-07:002019-03-12T11:45:30.213-07:00Watch out for NaNs in the slab test!
instead of &q...Watch out for NaNs in the slab test!<br />instead of "vec3 tmin = min(t0,t1), tmax=max(t0,t1)" do "vec3 tmin = min(t0,t1), tmax=max(t1,t0)". Switching the order of the operands makes sure you don't propagate unnecessary NaNs.Carmelo J Fernández-Agüera Tortosahttps://www.blogger.com/profile/13578762820521560567noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-85142332254388978362019-03-12T08:20:06.331-07:002019-03-12T08:20:06.331-07:00Also interesting: Ganestam et al.'s Bonsai con...Also interesting: Ganestam et al.'s Bonsai construction scheme is very fast and can be augmented with a very effective spatial split extension. It's a weird one, since it does the spatial splitting as a preprocess, but it yields nearly perfect trees (compared to straight Bonsai).Jacco Bikkerhttps://www.blogger.com/profile/02786431312546507699noreply@blogger.comtag:blogger.com,1999:blog-8350257063773144600.post-45719968286491223082019-03-12T08:14:21.815-07:002019-03-12T08:14:21.815-07:00Some additional things:
- Traverse the nodes front...Some additional things:<br />- Traverse the nodes front-to-back. That way, you can cull nodes of which the entry point is beyond the currently nearest intersection. Expect a very significant performance increase.<br />- Do not bother sorting nodes for shadow rays and do an early out.<br />- Have dedicated shadow ray code; they just need to store a boolean, not an intersection distance + u + v.<br />- Store your intersection result as a set of four floats using a single float4 write. Modern GPUs will do a single mem op for this, which is much faster than e.g. writing 2 or 3 separate floats. Same for reading back that data later.<br />- Improve SAH construction speed with some binning. Yields nearly identical trees at a fraction of the cost.<br />- Consider spatial splits. A lot of extra work, but you can expect another 20% efficiency improvement for ray queries.<br />- Use a top-level BVH over BVhs built per scene graph node. This lets you construct static geometry BVHs offline, while dynamic objects move for free if the motion can be expressed with a matrix, or with a low quality BVH if it can't.<br /><br />There's probably more, but this should be a good start.<br /><br />- Jacco.Jacco Bikkerhttps://www.blogger.com/profile/02786431312546507699noreply@blogger.com