Partitioned convolution
Splitting a long impulse response into blocks and convolving each one via FFT, so convolution reverb becomes both affordable and close to latency-free.
Direct convolution costs one multiply per impulse-response sample for every output sample – for a three-second cathedral at 48 kHz, around 144,000 per sample. FFT convolution is vastly cheaper but has to wait for a full block, and a block as long as the impulse response means seconds of latency.
The fix is to chop the impulse response up. With equal-sized blocks you get one block of latency and can reuse spectra computed for earlier blocks. Bill Gardner showed in 1995 how to get all the way to zero: convolve the very start of the response directly as an FIR filter, then double the block size for each step further back. Everything from Altiverb to the browser's ConvolverNode runs on some variant of that.