new #ptx instructions from cuda sdk 13.4: redplait.blogspot.com/2026/1… 21 new instructions 17 new attributes 127 new forms

extracted rt kernels from #cuda 13.4: github.com/redplait/denvdis/… no, this still not fixed (file name sm236.elf): xitter.cf/real_redp/status/21121… sm107 version of this file in sm175.elf

inefficient uniform registers allocation in #ptxas: redplait.blogspot.com/2026/1… seems that ptxas has bug in its register allocator and marks uniform registers for desc operand as dirty - and so allocating and assign new pair every time

comparison of #nvcc vs #clang #nvptx backend, part 3: redplait.blogspot.com/2026/1… dirty hacks to find added cl options & passes idc & perl of course

comparison of #nvcc vs clang #nvptx backend, part 2: redplait.blogspot.com/2026/1… - cicc does not have MLIR - cicc has 870 builtins vs 715 in nvptx - cicc has 3183 intrinsics vs 1998 in nvptx

mwa-ha-ha: amazon.com/Vibe-Coding-Done-… 312 pages I can save you 50$ and several days for reading: just avoid it like the plague

nvvm instrinsics from clang: github.com/redplait/denvdis/… - 1998 from cicc: github.com/redplait/denvdis/… - 4053 that's just completely another level of nvptx backend, right?

comparison of #prx/#sass code from nvcc vs clang: redplait.blogspot.com/2026/1… - nvcc code 15–18% faster - bigger and more amenable to stall counts reducing - all mlir based compilers use vanilla llvm nvptx backend :-(

#ptx generated by #LLVM #NVPTX: redplait.blogspot.com/2026/1… nvptx uses 58.9% of all ptx instructions cicc 57.8% missed in clang: _ldsm _mma discard dp2a hi/lo madc hi/lo p2r trap

read "the anatomy of go" very good book about mediocre PL: redplait.blogspot.com/2026/0… @func25: could you check whether I lied anywhere?

my tool for #cuda #cubin extracting/patching now support x86-64 object files: redplait.blogspot.com/2026/0…

#ptx asm in #llvm ir: redplait.blogspot.com/2026/0… llvm doc lies about operand constraint codes you can swap operand and this makes parsing/comparison of instructions non-trivial nvidia's llvm backend cicc uses ~5-7% of instructions with undocumented attributes not avail in clang

shrinking holes in #SASS allocated registers: redplait.blogspot.com/2026/0… due to using of wide operations you can shrink only small amount of holes - in average 1-2% of functions

add to my #cubin editor cool command to replace regs/predicates inside instruction, like 398 R P1 P2 and ISETP.GE.AND P1,PT,R07,RZ,PT ?trans1 will become ISETP.GE.AND P2,PT,R07,RZ,PT ?trans1 github.com/redplait/denvdis/…

chilling article literally describing reasons behind the emergence of Skynet: xitter.cf/cdleary/status/2094878… "we don’t really need to understand what the kernel does line by line" We are doomed

in addition to xitter.cf/real_redp/status/20944… I add some stat to estimate potential reg count decrease 1.7-2.7% for gpr 0.1 for uniform registers (without UR0-UR3)

Cyclomatic complexity of #SASS code: redplait.blogspot.com/2026/0… dead(ABI-Forced?) register writes Register Allocation Holes - probably PTXAS ran some optimization passes after AllocateRegisters and they eliminated corresponding instructions. Or it's just bug

funny - every day I mute 5-6 persons with senseless "how amazing modern AI is/coding is solved" on other hand It is crystal clear that standard of modern software has hit rock bottom—down in the basement It seems there is direct cause-and-effect relationship between these facts

that's not joke: Functional Programming in #PHP amazon.com/Functional-Progra…

being a pathological fan of #perl of course I made binding for xitter.cf/real_redp/status/20904… enjoy: github.com/redplait/denvdis/…