Real-time software rendering with AVX-512

Описание: Impressive demo showcasing modern software rendering techniques performed entirely on the CPU using AVX-512 intrinsics. Written in C++ and compiled using Microsoft Visual Studio 2022.

This particular example demonstrates a popular new technique called ray marching, which allows us to render abstract shapes with unlimited resolution and detail, which does not use traditional polygonal rendering primitives.

Key characteristics:
Multithreaded
Tile-based
SSE4, AVX2, and AVX-512
Vanilla x86, too!
Texture mapping with bilinear and trilinear filtering.
Checkerboard rendering
Raymarching
Written in C++ using SIMD intrinsics

A quick note about CPU occupancy:
CPU utilization or 'occupancy' is low at high FPS because because uploading the final image to the display buffer over PCI-e is slower than rendering the entire image. For example, it may only take 2 milliseconds to render the spinning cube on desert landscape, but it can take maybe 5 or 6 milliseconds to upload it. This synchronization causes the CPU to idle between frames and is why the effective CPU clock speed is under 1GHz in some cases.

For more complex scenes where rendering can take 100 milliseconds, the idle time ratio is much lower therefor CPU occupancy is closer to peak performance but there are still a few milliseconds to account for.

The FPS shown in the frame is the CPU-only frame-time, and does not account GPU upload or V-Sync.

A future optimization would be to double-buffer the output buffer and concurrently render the next frame while uploading the prior frame to GPU.


To provide a frame of reference, my GeForce RTX 3080 is able render "Abstract Terrain" at around 120 FPS, or around 4 or 5x faster, but that's without 'checkerboard' rendering. By these estimates, I would estimate the 3080 is approximately 8 to 10 times faster than Sapphire Rapids overall.