Serial C++ can be parallel too

In this talk, we will see how serial C++ can utilize hardware parallelism. We will begin with SIMD (Single Instruction, Multiple Data), where one instruction processes a whole batch of values at once. The path will run from hand-written vectorized assembly, to compiler intrinsics, to the new C++26 std::simd feature, and finally to auto-parallelization, with examples of what works, what doesn’t, and how to verify the compiler’s choices. We will see why SIMD shines in structured compute-bound code, and why it sometimes falls short.

We will then broaden the scope to memory-level parallelism (MLP) - a hardware feature that overlaps multiple independent memory accesses to reduce overall latency. We will discuss how to take advantage of MLP in C++ through bulk operations and APIs, and show how this approach can bring substantial benefits even in less structured code with irregular memory access patterns.