4 ms·
Honest question, I feel like kernels are usually short enough that you can fully understand their performance in the development cycle before you even deploy th
by saagarjha 4mo ago
Honest question, I feel like kernels are usually short enough that you can fully understand their performance in the development cycle before you even deploy them. If you get different results in production this seems to me that you didn’t spend enough time understanding what’s going on earlier. Are there things you genuinely can’t get from this workflow?
- SyzygyRhythm 4mo agoSometimes you have to optimize other people's code. Also, sometimes code behaves unexpectedly depending on the data, say over a certain size threshold. And sometimes it behaves differently on different hardware. You don't always find these things out until production.
- saagarjha 4mo agoI don't really understand how PC sampling is going to help you with this
- gnurizen 4mo agoA lot of times the performance is data dependent, like randomized data can result in poor warp utilization where as sorted data can be much more efficient because you have less warps sitting around idle. We have an upcoming blog post on decompression kernels that show this clearly, stay tuned!