On Friday, 18 October 2019 at 01:37:01 UTC, Walter Bright wrote:
Slides: https://digitalmars.com/articles/hits.pdf
Tangent time. In regards to floating point:
Unable to convince people that more precision is worthwhile
I'm actually waiting for quad floats to have hardware support. Registers are already wide enough, just nothing internally or in terms of instruction set for them yet.
But yeah. The reality is that the hardware I operate on uses 32-bit floats almost exclusively on the CPU (the sole exception I can think of is the main simulation timer, you don't want that as a 32-bit float). GPUs and shaders use 16-bit half floats extensively these days. But 64-bit is a rarity because operations take 40% more execution time on average with my own tests and we can forsake the accuracy for execution time.
This might change in a 4K TV world. Haven't really done too much with them yet (despite Quantum Break supporting 4K, but I didn't really notice anything glitchy that I could associate with floating point imprecision).
Still. 64- and 128-bit floats are quite useful for offline calculations. Make your data as accurate as possible, then let the runtime code use the fastest execution path it can.
