oof retweeted
HOLY. I just GAINZed the SHIT out of transforming by 4x4 rotation matrices, given Euler angle triplets (technically intrinsic Tait-Bryan angles, for all you math nerds)... This is for SH4ZAM, my SH4 assembly optimized fast math library for the Sega Dreamcast.
I had previously been applying these types of rotations by simply concatenating rotation matrices about the principal axes (X, Y, Z) in different orders, with one for every angle of the Euler triplet... and this was still so fast that I didn't have much incentive to gain further for quite some time.
...but I finally realized that I could actually optimize these types of rotations further, by creating single, dedicated routines, which operate entirely within the XMTRX 4x4 back-bank of floating-point registers directly, rather than round-tripping to and from memory, potentially paying a dcache stall penalty due to register spillage should pushing onto the stack cause a cache miss...
So for the price of more FPU work, and a few more total instructions, I'm basically eliminating virtually all intermediate memory accesses, by performing all of the work directly within register-space... but this price is worth it, because the extra instructions are all dual-dispatching superscalar pairs that pipeline extremely nicely...
So what are the GAINZ? This routine is now 35% faster than concatenating individual transform matrices about the X + Y + Z axes to perform the same operation... while it's a whole-ass 305-474% faster than the fastest you can do in plain C with the CGLM library! 💪
PS: OMG OMG OMG what TF is the screenshot? ...no comment. Helping Lobotomy out with his latest epic surprise port, which is a certain fighting game starring big-N characters from a certain purple cube console...
Footage coming soon of it running on the Dreamcast, once it's not running like dog shit and running out of VRAM constantly! 🤣
oof retweeted
The first thing I learned in engineering school is that a good developer is someone who knows how to do a good Google search to find the piece of code they need and copy paste it
oof retweeted
Crossworlds has saved sonic team's rep by showing if they are allowed to they can make new models and have more quality renders. Them also speaking that they needed the dlcs for funding make these literally a cry for help 😭
What fucks me up here is that credits have explicit license reminders. I wonder why so out of place 🤔
And these "None, no" etc confirms even further that it's AI slop layered on top of AI slop layered on top of something human
github.com/DevZer0D4Y/SonicF…
23
oof retweeted
Working on the SH4 assembly implementations of some of the most hotly requested routines for the next release of my fast vector math library, SH4ZAM, for the Sega Dreamcast.
This time I’ll be adding support for billboarding rotation transforms to the XMTRX and in-memory 4x4 matrix APIs, allowing you to initialize and apply such transforms, fully utilizing the FPU and SIMD instructions faster than can be done in software using C or C++.
…yes, while almost every retro gamedev homebrew community has progressed towards not even actually writing code anymore, thanks to Claude, my cave man, Stone Age ass is still hand-writing ASM routines in a notebook… 🤣
