ok here it is. a free mini book on assembly, from bits all the way down to what your gpu actually runs.
risc-v > x86 > simd > ptx > sass. 129 pages.
not gonna lie, i worked very hard on this. i turned my transcripts of what i understood into this, then had ai refine those transcripts. but trust me, its worth the read.
every risc-v example was tested against a reference (245 checks), every compiler listing is what clang actually printed, and every benchmark was measured. the cuda kernels were compiled but not run (no gpu on me), and the book explicitly says that on the page.
my fav part: the same matmul, 3 loop orders, went from 1.8 > 68 gflop/s on one core. 38x faster without removing a single flop.
would appreciate your feedback, guys.
https://t.co/Ee0vWzoM4M