[avionicharshit-byte]

towards #561 , this does not close it , see the bottom

matmul.cpp already had a non-DOTPROD fallback , cactus_dotq_lane0..3 behind CACTUS_DOTQ_LANE , used in 52 places. but 28 other call sites in the same file still call vdotq_laneq_s32 directly , and attention_hybrid.cpp calls vdotq_s32 in 10 places with no guard at all. so on a target without dotprod the code does not fall back , it fails to compile.

what this does

moves the guard block into cactus-kernels/src/neon_dotprod.h so both files share one switch , adds CACTUS_DOTQ for the non lane form which had no fallback at all , and routes the 38 raw sites through the macros. nothing changes on a dotprod target , the macros expand to the same intrinsics.

testing , on an m-series mac

• -march=armv8.2-a+fp16 , no dotprod. before: 32 errors in matmul.cpp , 10 in attention_hybrid.cpp. after: 0 and 0
• bit exact , 200k random vectors through both paths across all 5 call forms , 0 mismatches. its integer maths so exact , not approximate
• ./test.sh passes 50/50 native , and 50/50 again with the #undef __ARM_FEATURE_DOTPROD line uncommented so every site is forced down the fallback

the loss

the fallback is slower , measured with that switch on:

matmul_cq1 1024^3         2.818ms -> 17.781ms
matmul_cq4 1024^3         3.182ms -> 18.179ms
matmul_cq1 1x1024x1024    0.028ms -> 0.047ms

so about 6x on large gemm and 1.7x on the gemv shape decode uses. its a correctness path , not a speed path , and it only runs where the fast one would have been an illegal instruction.

what this does not fix

pi 4 still will not build. the __fp16 arithmetic is a separate armv8.2 feature and has no fallback anywhere , with this branch applied its still 108 errors at -march=armv8-a across 11 files , all fullfp16. happy to do that next , either folded into this pr or as a second one.

one thing , i put the header at cactus-kernels/src/neon_dotprod.h , contributing says one header per folder so say the word and ill fold it into cactus_kernels.h instead.

[Udit8348]

For now, do not prioritize any work to bring up targets below ARM v8.2+
These processors are not targeted for "neural" workloads, in a way that cactus hw optimizations can take advantage of. (ie: fp16 support) There are obviously methods to emulate and add legacy processor support, similar to how PyTorch does. On the other hand, this is not priority in my opinion and just clouds the src size, testing support, etc.

For now, it seems like build is configured to fail on unsupported architectures (https://github.com/cactus-compute/cactus/pull/768) This is sufficient observability for future debugging. IMO its better to first gauge audience interest with some kind of questionnaire in the readme.

[avionicharshit-byte]

tested on an actual pi 4 , not cross compiled

pi 4 model b rev 1.4 , cortex-a72 , aarch64 , debian 13 , gcc 14.2. /proc/cpuinfo features are fp asimd evtstrm crc32 cpuid , so no asimddp and no asimdhp , a true armv8.0 target and the same board as #561

upstream main on this pi compiles fine with the current -march=armv8.2-a+fp16+simd+dotprod+i8mm and then dies at runtime. cactus-kernels/test.sh gives Illegal instruction on 6 of the 7 test binaries , 1 suite finishes , 13 of the 50 tests get to run at all

at an honest armv8.0 baseline , -march=armv8-a+simd , upstream does not compile , 118 errors across 11 files

with this branch , 88 errors , same 11 files , and every remaining one is __fp16. isolating the dotprod half with -march=armv8.2-a+fp16 , fp16 on and dotprod off:

                       upstream   this branch
matmul.cpp                   16             0
attention_hybrid.cpp         30             0

what i cannot show yet is the suite passing on the pi , it still wont build there until __fp16 arithmetic has a fallback too. thats the other half , 88 errors worth , same shape of work

@HenryNdubuaku on that , do you want it folded into this pr , a second one stacked on top , or would you rather not carry a slow path at all. either way this pr changes no build flags , so nothing moves for chips that have dotprod , and i have the pi here to measure whichever way you prefer