← All resources

IBL Optimization Study III: Shared memory and Latency Hiding

Auto-discovered

Blog Advanced

Carmelo J Fernández-Agüera Tortosa

Visit resource ↗

Description

IBL Optimization Study III: Shared memory and Latency Hiding is the third Technik blog post by Carmelo J Fernandez-Aguera Tortosa optimizing GGX specular prefiltering for image-based lighting.

NSight captures on RTX 5080 and RTX 3050 Ti show latency-bound texture waits. The post applies partial loop unrolling, group-shared Hammersley sequences, shared tangent-space GGX samples, tangent-space weight and LOD math, and shared LOD precompute, cutting 3050 Ti GGX time from about 2.17 ms to under 1 ms. Dead ends include fp16, larger groups, and CPU vectors in VRAM.

It is for advanced graphics programmers tuning online IBL prefilters who can read compute-shader and profiler detail. Prior series posts supply irradiance context.

The article is a free Blogspot page with timing tables, NSight notes, and commit references.

It is useful for turning a naive Karis-style GGX filter into a latency-aware shared-memory pass before async compute or adaptive sample counts.