Going Faster With Alignment

So, in the previous post I talked about, fast_remove_if, and going faster than the std::remove_if function. Now there is a further improvement that can be made to go *even* faster, but with no need to modify any functions – instead it involves modifying the alignment of the data in the vector.

Here’s the object we used, it has no specific alignment.

		class TestObject {
		public:
			bool remove{ false };
			int i{ 8 };
			float k{ 2 }, a{ 1 };
			double d{ 99.99 };

			int i2{ 81 };
			float k2{ 21 }, a2{ 11 };
			double d2{ 991.99 };

			int i3{ 80 };
			float k3{ 20 }, a3{ 10 };
			double d3{ 990.990 };
		};

To make this object faster to work with, it needs to be aligned to the stack size, which on x64 bit systems, is 16-bytes.

Here’s how we do that:

	__declspec(align(16)) class TestObject {
	public:
		bool remove{ false };
		int i{ 8 };
		float k{ 2 }, a{ 1 };
		double d{ 99.99 };

		int i2{ 81 };
		float k2{ 21 }, a2{ 11 };
		double d2{ 991.99 };

		int i3{ 80 };
		float k3{ 20 }, a3{ 10 };
		double d3{ 990.990 };
	};

And now, magically, our object is going to be speedier. Here’s some results from the previous unit test, using this new aligned object.

Given 19730 objects, with a randomly selected amount to remove.

Results for fast_remove_if:

• one percent of objects: 65 avg micro
• varying percent of objects: 116 avg micro

Results for erase, std::remove_if

• one percent of objects: 88 avg micro
• varying percent of objects: 117 avg micro

An additional improvement of around 20%.