// 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.

$assert BATCH_TILE % 4 == 0
$assert BATCH_TILE >= 4
$assert RR_STEPS in [1, 2]
$assert DIV_ALGO in ["div", "nr2fma", "nr2recps", "nr1recps1fma"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
$VMULSUBQ_F32 = "vfmsq_f32" if FMA else "vmlsq_f32"
#include <assert.h>

#include <arm_neon.h>

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


extern XNN_INTERNAL const float xnn_table_exp2_k_over_64[64];

void xnn_f32_sigmoid_ukernel__${"neonfma" if FMA else "neon"}_rr${RR_STEPS}_lut64_p2_${DIV_ALGO}_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const void* params)
{
  assert(n % sizeof(float) == 0);

  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p23f);
  // The largest z for which sigmoidf(-z) is normalized.
  // This number is also the largest z for which expf(-z) is normalized.
  const float32x4_t vdenorm_cutoff = vmovq_n_f32(0x1.5D589Ep+6f);
  const float32x4_t vminus_log2e_x64 = vmovq_n_f32(-0x1.715476p6f);
  $if RR_STEPS == 1:
    const float32x4_t vln2_o64 = vmovq_n_f32(0x1.62E43p-7f);
  $else:
    $if FMA:
      const float32x4_t vln2_o64_hi = vmovq_n_f32(0x1.62E43p-7f);
      const float32x4_t vln2_o64_lo = vmovq_n_f32(-0x1.05C61p-35f);
    $else:
      // Last 13 bits are zeroes
      const float32x4_t vln2_o64_hi = vmovq_n_f32(0x1.630000p-7f);
      const float32x4_t vln2_o64_lo = vmovq_n_f32(-0x1.BD0106p-19f);
  const float32x4_t vone = vmovq_n_f32(1.0f);

  const float32x4_t vc2 = vmovq_n_f32(0x1.FFFF0Ap-2f);

  const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));

  $if BATCH_TILE > 4:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;

      // General structure of the algorithm:
      //           / exp(x) / (1 + exp(x)) if x <= 0
      //   f[x] := 
      //           \ 1 - f[-x] if x >= 0
      //
      // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
      // then replace result with 1 - f[-z] if x >= 0.
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vz${ABC[N:N+4]} = vabsq_f32(vx${ABC[N:N+4]});

      // Compute reduced argument n := round(-z * 64 / log(2)).
      // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
      // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
      // The trick with adding large number is valid only within certain bounds (|z * 64 / log(2)| <= 2**22, i.e.
      // |z| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs x outside of [-87.336544, 17.328678]
      // (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result  for such inputs at the
      // very end of the algorithm.
      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vz${ABC[N:N+4]}, vminus_log2e_x64);

      // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that sigmoidf(-z) is
      // normalized, i.e. 0 <= z <= 87.33642. As n has 6 fractional bits, we split s == 2**(n / 64) =
      // = 2**e * 2**(n / 64 - e), where e := int(n / 64). We create s in two steps:
      // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
      //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
      // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
      //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
      //    and thus the adjusted exponent is not lower than -126.
      //
      // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
      $for N in range(0, BATCH_TILE, 4):
        const int32x4_t ve${ABC[N:N+4]} = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), vmovq_n_s32(INT32_C(0x3F))), 17);

      // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
      $for N in range(0, BATCH_TILE, 4):
        const uint64x2_t vidx${ABC[N:N+4]} = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), vindex_mask));

      $for N in range(0, BATCH_TILE, 4):
        const uint64_t vidx${ABC[N:N+2]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 0);
        const uint64_t vidx${ABC[N+2:N+4]} = vgetq_lane_u64(vidx${ABC[N:N+4]}, 1);
        float32x2_t vl${ABC[N:N+2]} = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx${ABC[N:N+2]}]);
        float32x2_t vl${ABC[N+2:N+4]} = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx${ABC[N+2:N+4]}]);

      $for N in range(0, BATCH_TILE, 4):
        vl${ABC[N:N+2]} = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx${ABC[N:N+2]} >> 32)], vl${ABC[N:N+2]}, 1);
        vl${ABC[N+2:N+4]} = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx${ABC[N+2:N+4]} >> 32)], vl${ABC[N+2:N+4]}, 1);
        const float32x4_t vl${ABC[N:N+4]} = vcombine_f32(vl${ABC[N:N+2]}, vl${ABC[N+2:N+4]});

      // Adjust exponent of the value l fetched from the table to get the final s value.
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl${ABC[N:N+4]}), ve${ABC[N:N+4]}));

      // Subtract the large number back to get the final n := round(-z * 64 / log(2)) as a floating-point number.
      $for N in range(0, BATCH_TILE, 4):
        vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

      // Compute reduced argument t := (z + n * log(2) / 64). Note that -t = -z - n * log(2) / 64.
      $if RR_STEPS == 1:
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vz${ABC[N:N+4]}, vn${ABC[N:N+4]}, vln2_o64);
      $else:
        // Use Cody-Waite range reduction method (note two constants to represent log(2) / 64) to improve accuracy.
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vz${ABC[N:N+4]}, vn${ABC[N:N+4]}, vln2_o64_hi);

        $for N in range(0, BATCH_TILE, 4):
          vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vln2_o64_lo);

      // Compute degree-2 polynomial approxiatmion for exp(-t) on [-log(2)/128, log(2)/128].
      //   P1(t) = 1 + t * (-1 + t * c2)
      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vp${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vc2);

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = ${VMULSUBQ_F32}(vt${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      // Reconstruct the exp(-z) value:
      //   f = s * (1 + t * (-1 + t * c2))
      //     = s * (1 - t + t * (t * c2))
      //     = s - s * (t - t * (t * c2))
      //     = s - s * p
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vy${ABC[N:N+4]} = ${VMULSUBQ_F32}(vs${ABC[N:N+4]}, vs${ABC[N:N+4]}, vp${ABC[N:N+4]});

      // Denominator of the sigmoid fraction: 1.0 + exp(-z)
      $for N in range(0, BATCH_TILE, 4):
        const float32x4_t vd${ABC[N:N+4]} = vaddq_f32(vy${ABC[N:N+4]}, vone);

      $if DIV_ALGO == "div":
        // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vf${ABC[N:N+4]} = vdivq_f32(vy${ABC[N:N+4]}, vd${ABC[N:N+4]});
      $else:
        // Use Newton-Raphson method (2 iterations) to compute reciprocal of denominator.
        // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
        // Thus the reciprocal of the denominator never overflows.
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vr${ABC[N:N+4]} = vrecpeq_f32(vd${ABC[N:N+4]});

        $if DIV_ALGO == "nr2fma":
          $for N in range(0, BATCH_TILE, 4):
            vr${ABC[N:N+4]} = vfmaq_f32(vr${ABC[N:N+4]}, vr${ABC[N:N+4]}, vfmsq_f32(vone, vr${ABC[N:N+4]}, vd${ABC[N:N+4]}));
        $else:
          $for N in range(0, BATCH_TILE, 4):
            vr${ABC[N:N+4]} = vmulq_f32(vr${ABC[N:N+4]}, vrecpsq_f32(vr${ABC[N:N+4]}, vd${ABC[N:N+4]}));

        $if DIV_ALGO == "nr2recps":
          $for N in range(0, BATCH_TILE, 4):
            vr${ABC[N:N+4]} = vmulq_f32(vr${ABC[N:N+4]}, vrecpsq_f32(vr${ABC[N:N+4]}, vd${ABC[N:N+4]}));
        $else:
          $for N in range(0, BATCH_TILE, 4):
            vr${ABC[N:N+4]} = vfmaq_f32(vr${ABC[N:N+4]}, vr${ABC[N:N+4]}, vfmsq_f32(vone, vr${ABC[N:N+4]}, vd${ABC[N:N+4]}));

        // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
        $for N in range(0, BATCH_TILE, 4):
          float32x4_t vf${ABC[N:N+4]} = vmulq_f32(vy${ABC[N:N+4]}, vr${ABC[N:N+4]});

      // For inputs below denormal cutoff, replace output with +0.0f.
      // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
      $for N in range(0, BATCH_TILE, 4):
        vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcagtq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

      // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
      $for N in range(0, BATCH_TILE, 4):
        const uint32x4_t vm${ABC[N:N+4]} = vcltq_f32(vx${ABC[N:N+4]}, vmovq_n_f32(0.0f));

      $for N in range(0, BATCH_TILE, 4):
        vf${ABC[N:N+4]} = vbslq_f32(vm${ABC[N:N+4]}, vf${ABC[N:N+4]}, vsubq_f32(vone, vf${ABC[N:N+4]}));

      $for N in range(0, BATCH_TILE, 4):
        vst1q_f32(y, vf${ABC[N:N+4]}); y += 4;
    }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    const float32x4_t vx = vld1q_f32(x); x += 4;

    // General structure of the algorithm:
    //           / exp(x) / (1 + exp(x)) if x <= 0
    //   f[x] := 
    //           \ 1 - f[-x] if x >= 0
    //
    // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
    // then replace result with 1 - f[-z] if x >= 0.
    const float32x4_t vz = vabsq_f32(vx);

    // Compute reduced argument n := round(-z * 64 / log(2)).
    // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
    // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
    // The trick with adding large number is valid only within certain bounds (|z * 64 / log(2)| <= 2**22, i.e.
    // |z| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs x outside of [-87.336544, 17.328678]
    // (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result  for such inputs at the
    // very end of the algorithm.
    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vminus_log2e_x64);

    // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that sigmoidf(-z) is
    // normalized, i.e. 0 <= z <= 87.33642. As n has 6 fractional bits, we split s == 2**(n / 64) =
    // = 2**e * 2**(n / 64 - e), where e := int(n / 64). We create s in two steps:
    // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
    //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
    // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
    //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
    //    and thus the adjusted exponent is not lower than -126.
    //
    // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
    const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);

    // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
    const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
    // Adjust exponent of the value l fetched from the table to get the final s value.
    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));

    // Subtract the large number back to get the final n := round(-z * 64 / log(2)) as a floating-point number.
    vn = vsubq_f32(vn, vmagic_bias);

    // Compute reduced argument t := (z + n * log(2) / 64). Note that -t = -z - n * log(2) / 64.
    $if RR_STEPS == 1:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vln2_o64);
    $else:
      // Use Cody-Waite range reduction method (note two constants to represent log(2) / 64) to improve accuracy.
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vln2_o64_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vln2_o64_lo);

    // Compute degree-2 polynomial approxiatmion for exp(-t) on [-log(2)/128, log(2)/128].
    //   P1(t) = 1 + t * (-1 + t * c2)
    float32x4_t vp = vmulq_f32(vt, vc2);
    vp = ${VMULSUBQ_F32}(vt, vp, vt);

    // Reconstruct the exp(-z) value:
    //   f = s * (1 + t * (-1 + t * c2))
    //     = s * (1 - t + t * (t * c2))
    //     = s - s * (t - t * (t * c2))
    //     = s - s * p
    const float32x4_t vy = ${VMULSUBQ_F32}(vs, vs, vp);

    // Denominator of the sigmoid fraction: 1.0 + exp(-z)
    const float32x4_t vd = vaddq_f32(vy, vone);

    $if DIV_ALGO == "div":
      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
      float32x4_t vf = vdivq_f32(vy, vd);
    $else:
      // Use Newton-Raphson method (2 iterations) to compute reciprocal of denominator.
      // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
      // Thus the reciprocal of the denominator never overflows.
      float32x4_t vr = vrecpeq_f32(vd);

      $if DIV_ALGO == "nr2fma":
        vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
      $else:
        vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));

      $if DIV_ALGO == "nr2recps":
        vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
      $else:
        vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));

      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
      float32x4_t vf = vmulq_f32(vy, vr);

    // For inputs below denormal cutoff, replace output with +0.0f.
    // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));

    // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
    const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
    vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));

    vst1q_f32(y, vf); y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    const float32x4_t vx = vld1q_f32(x);

    // General structure of the algorithm:
    //           / exp(x) / (1 + exp(x)) if x <= 0
    //   f[x] := 
    //           \ 1 - f[-x] if x >= 0
    //
    // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
    // then replace result with 1 - f[-z] if x >= 0.
    const float32x4_t vz = vabsq_f32(vx);

    // Compute reduced argument n := round(-z * 64 / log(2)).
    // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
    // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
    // The trick with adding large number is valid only within certain bounds (|z * 64 / log(2)| <= 2**22, i.e.
    // |z| <= 0x1.62E43p+15 = 45426.09375), but that is acceptable, because inputs x outside of [-87.336544, 17.328678]
    // (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result  for such inputs at the
    // very end of the algorithm.
    float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vz, vminus_log2e_x64);

    // Create a floating-point number s (scale) such that s := 2**(n / 64) for such inputs that sigmoidf(-z) is
    // normalized, i.e. 0 <= z <= 87.33642. As n has 6 fractional bits, we split s == 2**(n / 64) =
    // = 2**e * 2**(n / 64 - e), where e := int(n / 64). We create s in two steps:
    // 1. Fetch 2**(n / 64 - e) = 2**(n % 64) from the table using the 6 low bits of n, as integer. Note that the
    //    fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
    // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
    //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
    //    and thus the adjusted exponent is not lower than -126.
    //
    // Extract e from bits 6:14 of n and shift it into bits 23:31 (position of floating-point exponent).
    const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);

    // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
    const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
    const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
    const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
    float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
    // Adjust exponent of the value l fetched from the table to get the final s value.
    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));

    // Subtract the large number back to get the final n := round(-z * 64 / log(2)) as a floating-point number.
    vn = vsubq_f32(vn, vmagic_bias);

    // Compute reduced argument t := (z + n * log(2) / 64). Note that -t = -z - n * log(2) / 64.
    $if RR_STEPS == 1:
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vln2_o64);
    $else:
      // Use Cody-Waite range reduction method (note two constants to represent log(2) / 64) to improve accuracy.
      float32x4_t vt = ${VMULADDQ_F32}(vz, vn, vln2_o64_hi);
      vt = ${VMULADDQ_F32}(vt, vn, vln2_o64_lo);

    // Compute degree-2 polynomial approxiatmion for exp(-t) on [-log(2)/128, log(2)/128].
    //   P1(t) = 1 + t * (-1 + t * c2)
    float32x4_t vp = vmulq_f32(vt, vc2);
    vp = ${VMULSUBQ_F32}(vt, vp, vt);

    // Reconstruct the exp(-z) value:
    //   f = s * (1 + t * (-1 + t * c2))
    //     = s * (1 - t + t * (t * c2))
    //     = s - s * (t - t * (t * c2))
    //     = s - s * p
    const float32x4_t vy = ${VMULSUBQ_F32}(vs, vs, vp);

    // Denominator of the sigmoid fraction: 1.0 + exp(-z)
    const float32x4_t vd = vaddq_f32(vy, vone);

    $if DIV_ALGO == "div":
      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
      float32x4_t vf = vdivq_f32(vy, vd);
    $else:
      // Use Newton-Raphson method (2 iterations) to compute reciprocal of denominator.
      // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
      // Thus the reciprocal of the denominator never overflows.
      float32x4_t vr = vrecpeq_f32(vd);

      $if DIV_ALGO == "nr2fma":
        vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
      $else:
        vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));

      $if DIV_ALGO == "nr2recps":
        vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
      $else:
        vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));

      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
      float32x4_t vf = vmulq_f32(vy, vr);

    // For inputs below denormal cutoff, replace output with +0.0f.
    // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));

    // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
    const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
    vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));

    float32x2_t vf_lo = vget_low_f32(vf);
    if (n & (2 * sizeof(float))) {
      vst1_f32(y, vf_lo); y += 2;
      vf_lo = vget_high_f32(vf);
    }
    if (n & (1 * sizeof(float))) {
      vst1_lane_f32(y, vf_lo, 0);
    }
  }
}