stelliferous

Search uses release 0.0.2.0 · 2,086 definitions

Releases (2)

def gelu_tanh_ofPyTorch's gelu(approximate='tanh'): 0.5 x (1 + tanh(sqrt(2/pi) (x + 0.044715 x^3))).
def gelu_tanh_value
def identity
def relu
def relu6
def leaky_relu
def sigmoid
def silu
def tanh
def hard_sigmoid
def hard_swish
def gelu_tanh
def is_identityThe identity needs no pass at all.
def operationsOperations per element, excluding the pass that applies them.
def unknown_operationsA custom function (F.conv2d_custom, F.apply) states its scalar operations per element: one per arithmetic, comparison or select; F32.exp about 37 and F32.tanh about 82 with their conversions.

def nonzeroOne for a nonzero word, zero otherwise, without a Bool match.
def number
def ordered
def less_bitMagnitudes fit in 31 bits, so the top bit of their difference is the unsigned comparison.
def equal_bit
def mask_value
def less_words
def less_equal_words
def equal_words
def not_equal_words
def lessContext comes first in Operations.lanes; the tensor element comes second.
def less_equal
def greater
def greater_equal
def equal
def not_equal
type Bounds
def clamp
def less_elementsBinary element order is left then right.
def less_equal_elements
def greater_elements
def greater_equal_elements
def equal_elements
def not_equal_elements
def selectSelection only changes words; the chosen payload is never FP arithmetic.
def selected
def rejected
def union

def multiply_add
def weighted_add
def zip_mac
def exponential_polynomialThe Cephes expf polynomial after its constant and linear terms.
def from_bitsThe 32-bit word of a U32, read as an F32.
def exponential_remainderexp(r)-1-r near |r| <= ln(2)/2.
def logistic_denominatorb=RN(1+q).
def selected
def silu_finishAbove 18 SiLU rounds to x.
def logistic_scaled
def sigmoid_reducedu=-value, n approximates u/ln(2) to the nearest integer, r=u-n*ln(2).
def sigmoid_input
def sigmoidAt x>=18 sigmoid rounds to 1.
def activate
def tanh_finishSign restoration and the tiny interval are word selections.
def tanh_magnitudeAt |x|>=10, 1-tanh(|x|)<2*exp(-20)<2^-25.
def tanh_smallChebyshev interpolation of (tanh(sqrt(z))/sqrt(z)-1)/z, |t|<=0.75.
def tanh_remainderOne additional term for the expm1 correction in the tanh quotient.
def tanh_exponential
def tanh_input
def tanh
def exponential_scaleSplit 2^n at +64 for nonnegative inputs and -64 for negative inputs.
def exponential_preciseFastTwoSum retains the low bits of 1+r before adding the quadratic tail.
def exponential_reduced
def exponential_input
def exp
def logarithm_seriesFor s near r/(2+r), d = r-(2+r)*s gives the exact real identity log(1+r) = 2*atanh(s) + log(1+d/(1+s)).
def logarithm_normalizedThe exponent carry at mantissa word 0x3504f4 selects the same reduced interval as halving m above sqrt(2), without a floating-point selection.
def logarithm_finishScaling a subnormal by 2^23 is exact.
def logarithm_input
def log