src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut2048-p1-div-x8.c

// Auto-generated file. Do not edit!
//   Template: src/f32-vsigmoid/neon-lut2048-p1.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 <arm_neon.h>

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


extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_2048[2048];

void xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8(
    size_t n,
    const float* x,
    float* y,
    const void* params) XNN_OOB_READS
{
  assert(n % sizeof(float) == 0);

  const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p12f);
  const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p0f);
  const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
  const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
  const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFFEp-1f);
  const float32x4_t vone = vmovq_n_f32(1.0f);
  const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);

  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    const float32x4_t vx0123 = vld1q_f32(x); x += 4;
    const float32x4_t vx4567 = vld1q_f32(x); x += 4;

    const float32x4_t vz0123 = vabsq_f32(vx0123);
    const float32x4_t vz4567 = vabsq_f32(vx4567);

    float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
    float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);

    const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 12);
    const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 12);

    const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
    const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));

    const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
    const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
    float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx01]);
    float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx23]);
    const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
    const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
    float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx45]);
    float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx67]);

    vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx01 >> 32)], vl01, 1);
    vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx23 >> 32)], vl23, 1);
    const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
    vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx45 >> 32)], vl45, 1);
    vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx67 >> 32)], vl67, 1);
    const float32x4_t vl4567 = vcombine_f32(vl45, vl67);

    const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
    const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));

    vn0123 = vsubq_f32(vn0123, vmagic_bias);
    vn4567 = vsubq_f32(vn4567, vmagic_bias);

    float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
    float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);

    const float32x4_t vp0123 = vmulq_f32(vt0123, vc1);
    const float32x4_t vp4567 = vmulq_f32(vt4567, vc1);

    const float32x4_t vy0123 = vfmaq_f32(vs0123, vs0123, vp0123);
    const float32x4_t vy4567 = vfmaq_f32(vs4567, vs4567, vp4567);

    const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
    const float32x4_t vd4567 = vaddq_f32(vy4567, vone);

    float32x4_t vf0123 = vdivq_f32(vy0123, vd0123);
    float32x4_t vf4567 = vdivq_f32(vy4567, vd4567);

    vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
    vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));

    const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
    const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));

    vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
    vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));

    vst1q_f32(y, vf0123); y += 4;
    vst1q_f32(y, vf4567); y += 4;
  }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    const float32x4_t vx = vld1q_f32(x); x += 4;

    const float32x4_t vz = vabsq_f32(vx);

    float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
    const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);

    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_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);

    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
    vn = vsubq_f32(vn, vmagic_bias);
    float32x4_t vt = vfmaq_f32(vz, vn, vln2);

    const float32x4_t vp = vmulq_f32(vt, vc1);

    const float32x4_t vy = vfmaq_f32(vs, vs, vp);
    const float32x4_t vd = vaddq_f32(vy, vone);

    float32x4_t vf = vdivq_f32(vy, vd);
    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
    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);

    const float32x4_t vz = vabsq_f32(vx);

    float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
    const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);

    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_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
    float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
    vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
    vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
    const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);

    const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
    vn = vsubq_f32(vn, vmagic_bias);
    float32x4_t vt = vfmaq_f32(vz, vn, vln2);

    const float32x4_t vp = vmulq_f32(vt, vc1);

    const float32x4_t vy = vfmaq_f32(vs, vs, vp);
    const float32x4_t vd = vaddq_f32(vy, vone);

    float32x4_t vf = vdivq_f32(vy, vd);
    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
    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);
    }
  }
}