Closed ashvardanian closed 2 months ago
Major additions:
Minor fixes:
f16
i8
b8
bf16
Some crazy findings:
rsqrt
double
float
ValueError
OverflowError
math.sqrt
numpy.sqrt
sqrt
libc
Major additions:
Minor fixes:
f16
,i8
,b8
to Python buffers in 33f1b13bf16
to Rust, thanks to @WyctusSome crazy findings:
rsqrt
precision on Arm is not documented at allrsqrt
approximation fordouble
in AVX-512 is only 6x more accurate, than forfloat
ValueError
instead ofOverflowError
math.sqrt
overnumpy.sqrt
when dealing with NumPy arraysqrt
inlibc
is bit-precise