Loading feed…
Aras Pranckevičius0 Hot0 comments11 min read
Unroll a loop, make shader 10x faster · Aras' website
A GPU compute shader implementing Blender's anisotropic Kuwahara image filter was found to run dramatically slower than expected on Apple Silicon GPUs. Investigation using Xcode's shader profiler and a reverse-engineered Apple GPU disassembler revealed the shader compiler failed to unroll an inner loop, instead emitting hundreds of integer compare-select instructions per iteration to simulate array indexing in registers (since the GPU cannot index registers directly). Manually unrolling the loop with Blender's shading language's [[unroll]] attribute made the shader 10x faster on an M4 Max GPU (220ms to 20.4ms) and 2x faster on an RTX 3080Ti (17.8ms to 9.5ms), also reducing register usage and increasing occupancy. The fix has been merged into Blender via PR 164617 for the upcoming 5.3 release.
Source: https://aras-p.info/blog/2026/10/05/Unroll-a-loop-make-shader-10x-faster. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.
Recommended for you
Bình luận
Đang tải bình luận…
