[LAD] Re: Tape emulation - is DSP fundamentally flawed?

On Wed, Oct 16, 2024 at 03:04:39PM +0000, Juan P C wrote: I think you are mistaken in your understanding of how CPUs/DSPs function. If branch prediction etc produced different answers to mathematical operations, then nothing on computers would work with our current models. ¯\_(ツ)_/¯ It is absolutely true that the amount of time a series of operations takes CAN be variable due to (as you say) prediction, as well as effects of the memory subsystem (caches, bus interference), or (even bigger, since we're usually talking about desktop compute) CPU sharing. However, much as the telecom, aerospace, etc industries (that rely on deterministic compute) have found, you CAN have prediction and cache enabled and produce certified products (and even CPU sharing!), because you can typically show that they meet their compute deadlines on a well-designed system. But, if you're not happy with those technologies, you can often disable prediction, cache, etc on a modern processor and live with a more predictable (and predictably slower) experience. I don't recommend that, though it's something I've often used to isolate races and misconfigured memory at the kernel level. I think you are forgetting that a world without cache and speculative execution in CPUs is where we came from, and that there is a reason that we have moved on from that. You think DSPs aren't using the same predictive tricks as general-purpose CPUs to accelerate their performance? :) YES! I do agree with you, that plugins - if implemented properly - sound almost the same as analog! Algorithms don't care what timeline they are executed in nor how many times they are context switched. This is getting weird. N samples arrive, and N samples are either processed within their allotted time or not. If they are processed in time, then they are output on time. If they are not, then it will be heard as a glitch/distortion in the output, and the system is considered overloaded. You can build a custom audio system that runs without interrupts, but you'll be able to handle far lower audio bandwidth. Still, pure time-sliced systems ARE sometimes used for very critical systems, because you know they will always behave incredibly predictably. As with the other modern conveniences, though, such designs have been largely phased out. FWIW, you should probably expects to see 10ks of interrupts/sec on a general purpose CPU, not 10 million, to give you a sense of scale. If by 'problem' you mean 'input to output latency', then yes, that will sound better. I can't imagine small buffer sizes making the output sound better in any other way. -greg