@nprofile1q... also the overall filter has two command buffers in it and there's OS jitter affecting the cpu side round trip path

@nprofile1q... @nprofile1q... i want an oscilloscope with a nvlink interface one day. I can dream...

@nprofile1q... there's some jitter because the filter graph, rendering, and waveform capture are async to each other. These are single point numbers although i also have rough averages for a continuous run.

Also possible the different memory access patterns led to better cache hit rates in other shaders

@nprofile1q... @nprofile1q... my problem is getting data in and out the narrow channel that is VRAM lol. Almost all of my shaders are saturating the memory subsystem so the only major optimizations possible are to do less loads/stores or order them more efficiently for better coalescing or cache hit rates.

Which is why I'm beginning to look at kernel fusion for things like FFT + windowing so I can keep temporaries in registers and never touch the memory subsystem at all

And sure enough the other big shader in the same filter had a huge optimization possible too, 3.75 -> 0.78 ms for the shader and 8.8 -> 6.8 for the whole filter.

Not too bad considering it took 11.2 on the same workload a few hours ago, that's a 65% speedup overall

@nprofile1q... Not familiar with the internals of database engines to the point I could comment. They work best on massively data-parallel problems with predictable, linear memory address ordering.

In general the best case scenario is every thread making 32-bit reads and writes to consecutive word addresses, then doing a ton of flops before the next memory access. The further you diverge from this, the further from theoretical peak performance you get.

It feels like every time I look at something in ngscopeclient and go "hmm, this is too slow" and bang on it for a few hours I get a significant speedup out of somewhere.

@nprofile1q... Investigating... This is my first time using eventbrite so it's possible I messed something up.

@nprofile1q... there are so many confusing choices in what gets an emoji and what doesnt.

For example, silverware. We have knife, knife and fork, knife fork and plate, and spoon.

There is no standalone fork or plate emoji.

@nprofile1q... i spotted the fed and I'm not even in vegas

@nprofile1q... the one nice thing about the inefficiency of modern software is that if you have legitimately enormous compute requirements, teraflops are cheap if you actually put in the legwork to use them