We’re expanding our partnership with Singapore to help safely deploy AI at scale. 🇸🇬
Together with country experts, our new programs will focus on accelerating scientific discovery, advancing pandemic preparedness, and improving healthcare.
Find out more → https://t.co/V0rFGHXOGy
Qwen3.6 is under test in my home setup working predominantly in C++ codebases, and it’s more efficient than qwen3.5 models. I’ve tested using Haskell codebases and it works relatively well. It has become quite the assistant 🙂
definitely vectorised operations in the IR, ASM below maps to AArch64, macOS
.build_version macos, 15, 0
.globl"<encoded name of r>"
"<encoded name of r>":
mul.4hv0, v1, v0
ret
its instructive to learn that, I can dive even deeper but I'll let it go now.
Started going back to my first love, high performance computing and exploring Mojo.
fn r[size: Int](va: SIMD[DType.int8, size],
vb: SIMD[DType.int8, size],
out vc: SIMD[DType.int8, size]):
vc = va * vb
print("--- Unoptimized LLVM IR ---")
print(compile.compile_info[r[4], emission_kind="llvm"]())
print("\n--- Optimized LLVM IR ---")
print(compile.compile_info[r[4], emission_kind="llvm-opt"]())
which gives me
def r {
%3 = mul <4 x i8> %1, %0
ret <4 x i8> %3
}
@oxbow_lakes@debasishg I've often wonder that candidates probably know more than 1 way to solve the presented problems, but their brains might go "let me try something new to impress and increase my chances of getting an offer" - i know i've given that kind of thought more than once
@ktosopl A lot of people were misled unfortunately and I heard the so called victim was hired into Scala Center iirc. Anyway, scala will always be the language I loved.