Memory Management#

Using the C++ standard library brought the following problems:

• High overhead from thousands to tens of thousands of malloc/free calls per frame
• Cache-line-exclusive cost (x86 LOCK) on every smart-pointer refcount update
• Performance variance from many failure points plus OS scheduler intervention

So the decision was to write a custom memory data structure from scratch.

Built a page-based Arena allocator with Windows VirtualAlloc() — reserving memory a page at a time and handing back addresses with nothing but stack-pointer movement.

Objects whose lifetime is exactly one frame are grouped together, and the allocator is reset wholesale at the start of every frame — constant-folding N calls → 1 allocation/deallocation call.

Implemented a Free List/Pool to reuse previously allocated blocks, cutting allocation/deallocation overhead.

Both rendering data processing and object management improved over malloc/free:

Rendering at Scale#

Starting from 100v100 and scaling battle size further, frame drops appeared.

Root-caused it with CPU and GPU profilers, then used a graphics debugger to confirm the geometry shader stage of the Shadow Pass had poor parallelism. Switched to plain instanced rendering, improving it 15ms → 9ms.

Mapped the GPU memory used for matrix updates on the CPU side for the lifetime of the game, removing Map/Unmap and fence-sync cost every frame. Compressed 4x4 → 3x4 matrices for upload, improving bandwidth too.

CPU animation updates have no cross-unit dependency, so they run through parallel for, minimizing synchronization cost too.

template <typename F>
void ParallelFor(Thread_Group& group, int64_t count, F&& func);

Achieved 60 FPS at 1000v1000.

Path Finding#

Path finding, essential to an RTS, needed a triangulation algorithm underneath it. Considered an open-source library, but ran into:

• No deletion operation, needed to punch out destroyed obstacles
• Heavy codebases built mainly for things like geological simulation

So the geometry library was written from scratch.

Implemented the algorithm as a pipeline:

1. Constrained Delaunay Triangulation (CDT) generates the navmesh.
2. A* computes the route.
3. Simple Stupid Funnel simplifies it into a walkable path.

Faster build and runtime than well-known alternatives, plus deletion support:

Open-sourced the code on GitHub; picked up 100+ stars.