Tootfinder

There's still room to tune - shallow memory throughput is definitely suboptimal due to latency issues I need to chase - but with deep memory ngscopeclient ThunderScope is getting some pretty impressive performance.
2 channels @ 50M point memory depth (100M points per trigger) streaming at 7.5 WFM/s over LAN from across the building through a router. 40Gbase-SR4 from client to core switch and from core switch to router, then back to core switch, then 10Gbase-SR to the machine host…

Another evening, another few filters getting OOM speedups.
Tonight it was invert (27.3x, trivial memory bound shader that just outputs negative x[i] for each output sample) and the 8B/10B decode (didn't even bother to GPU it, just removing the redundant sampling operation by using the new CDR recovered-data output was enough for a 12.1x speedup and 6ms on my 50M point benchmark is fast enough I'm in no hurry to GPU... but at faster data rates with lower oversampling factors it …

Starting to look at some performance improvements for various scopehal filters as I begin the push for v0.2.
The benchmark dataset is P/N legs of 100baseTX Ethernet sampled at 500 Msps with a ThunderScope.
At 50M point memory depth, each waveform is 100 MB in size (2 channels * 8 bits * 500 Msps) and covers 100ms of real time.
The filter graph currently takes just shy of a second to run, meaning we're at 10% of real time. I want to improve on this.