xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-scalar-rr2-p5-div-x1.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f32-vsigmoid/scalar-rr2-p5-div.c.in
//   Generator: tools/xngen
//
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>
#include <math.h>

#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/vunary.h>


void xnn_f32_vsigmoid_ukernel__scalar_rr2_p5_div_x1(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n % sizeof(float) == 0);

  const float vmagic_bias = params->scalar_rr2_p5.magic_bias;
  const float vminus_log2e = params->scalar_rr2_p5.minus_log2e;
  const float vln2_hi = params->scalar_rr2_p5.ln2_hi;
  const float vln2_lo = params->scalar_rr2_p5.ln2_lo;
  const float vc5 = params->scalar_rr2_p5.c5;
  const float vc4 = params->scalar_rr2_p5.c4;
  const float vc3 = params->scalar_rr2_p5.c3;
  const float vc2 = params->scalar_rr2_p5.c2;
  const float vc1 = params->scalar_rr2_p5.c1;
  const float vone = params->scalar_rr2_p5.one;
  const float vdenorm_cutoff = params->scalar_rr2_p5.denorm_cutoff;

  do {
    const float vx = *x++;

    const float vz = fabsf(vx);

    float vn = vz * vminus_log2e + vmagic_bias;
    const float vs = uint32_as_float(float_as_uint32(vn) << 23);
    vn -= vmagic_bias;

    float vt = vn * vln2_hi + vz;
    vt = vn * vln2_lo + vt;

    float vp = vt * vc5 + vc4;
    vp = vt * vp + vc3;
    vp = vt * vp + vc2;
    vp = vt * vp + vc1;

    vt *= vs;
    const float ve = vt * vp + vs;
    const float vd = ve + vone;

    float vf = ve / vd;
    if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
      vf = 0.0f;
    }
    if XNN_UNPREDICTABLE(vx > 0.0f) {
      vf = vone - vf;
    }

    *y++ = vf;

    n -= sizeof(float);
  } while (n != 0);
}