src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c

// 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 <stddef.h>

#include <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/math-stubs.h>


// Table of exp2(k / 2048) values decremented (as integer) by (k << 12), k = 0..2048
extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_2048[2048];

void xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr2fma(
    size_t n,
    const float* input,
    float* output)
{
  assert(n % (4 * sizeof(float)) == 0);

  // Large number such that ulp(magic bias) == exp2(-11)
  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p12f);
  const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p0f);
  // Mask for the lowest 11 bits
  const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
  const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
  // Coefficient of polynomial approximation of exp(-t) ~ 1 + t * c1 on [-log(2)/2048, log(2)/2048]
  const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFFEp-1f);
  const float32x4_t vone = vmovq_n_f32(1.0f);
  // 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);

  for (; n != 0; n -= 4 * sizeof(float)) {
    const float32x4_t vx = vld1q_f32(input); input += 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 / log(2), 11).
    // We do it by adding a large number (magic bias), which cause rounding of the result to integer, then subtracing
    // the large number back. The trick with adding large number is valid only within certain bounds
    // (|-z / log(2)| <= 2**11, i.e. |z| <= 0x1.62E43p+10 = 1419.5654296875), 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 = vfmaq_f32(vmagic_bias, vz, vminus_log2e);

    // Create a floating-point number s (scale) such that s := 2**n for such inputs that sigmoidf(-z) is normalized,
    // i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**n = 2**int(n) * 2**frac(n). We create s
    // in two steps:
    // 1. Fetch 2**frac(n) from the table using the 11 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 int(n) 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 <= int(n) <= 0, and thus the adjusted exponent is not lower than -126.
    //
    // Shift bits 11:19 into 23:31 (position of floating-point exponent).
    const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);

    // Use bits 0:11 of n, as integer, as an index for table lookup of l := 2**frac(n).
    const uint64x2_t vidx = vreinterpretq_u64_s32(vshlq_n_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask), 2));
    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((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_2048 + (uint32_t) vidx_lo));
    float32x2_t vl_hi = vld1_dup_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_2048 + (uint32_t) vidx_hi));
    vl_lo = vld1_lane_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_2048 + (uint32_t) (vidx_lo >> 32)), vl_lo, 1);
    vl_hi = vld1_lane_f32((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_2048 + (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 / log(2), 11) as a floating-point number.
    vn = vsubq_f32(vn, vmagic_bias);

    // Compute reduced argument t := (z + n * log(2)). Note that -t = -z - n * log(2).
    float32x4_t vt = vfmaq_f32(vz, vn, vln2);

    // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/2048, log(2)/2048]:
    //   P(t) = 1 + t * c1 = 1 + p
    const float32x4_t vp = vmulq_f32(vt, vc1);

    // Reconstruct the exp(-z) value:
    //   e = s * (1 + t * c1)
    //     = s * (1 + p)
    //     = s + s * p
    const float32x4_t vy = vfmaq_f32(vs, vs, vp);

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

    // 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);
    vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
    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(output, vf); output += 4;
  }
}