Cargas vectorizadas con float4

A normal load fetches one value at a time. A vectorized load grabs several adjacent values in a single wider instruction, doing more work per request.

CUDA ships built-in vector types. A float4 packs four floats into one 16-byte bundle you can load or store together.

Reading a float4 from memory pulls four floats in a single transaction. That cuts the number of memory instructions a thread must issue by four.

To use vector loads on a plain float array, you cast the pointer. reinterpret_cast lets you read the same bytes as float4 elements.

A float4 must sit on a 16-byte boundary. If the data is not properly aligned, the load is illegal and your kernel can crash or read garbage.

Good news: buffers from cudaMalloc are aligned to at least 256 bytes. So freshly allocated device arrays are safe to read as float4 from the start.

Wider loads mean fewer outstanding requests for the same data. That can raise achieved bandwidth when a kernel is limited by memory, not math.

With float4 each thread now covers four elements. Your global index steps through float4 slots, so the total thread count drops by four.

float4 is not the only choice. Types like float2 and int4 give you 8- or 16-byte loads, so you can pick a width that fits your data.

If your array length is not a multiple of four, handle the extra remainder elements with a normal scalar loop after the vectorized part.

A float4 holds four values, so it uses more registers per thread. As always, confirm the bandwidth gain outweighs any drop in occupancy.

You reinterpret a float array as float4 but the kernel crashes. Why?

You learned that float4 loads four floats at once, cutting instructions and lifting bandwidth. Keep data aligned and handle leftover elements. 📦