I'm not saying that you can't do performance work without having done that. Just that you'll be at a disadvantage since you're at the mercy of whatever your HW vendor decides to disclose to you.
If you know this stuff you can work back from from first principals. With a high level memory architecture of a system(say tiled vs direct rendering GPU) you can reason about how certain operations will be fast and will be slow.
And your response was absurd. You don't rebuild a transmission in order to run a shop. You don't even rebuild a transmission as an engineer creating cars, you shop that out to an organization specializing in the extremely difficult task of designing and building transmissions. I wanted to avoid this waste of time, but here we are.
As for the rest of your comment about reasoning about performance, none of that requires the work you did. Again, neat project (for you), but completely unnecessary in general.
The primary reason for dropping down to native is to get better performance. If you're going to do that you'll leave 10x-50x performance on the table if you don't understand cache misses, prefetching and the other things that manual memory placement open up.