xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/neonfma-rr1-p5-x8-acc2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f32-raddstoreexpminusmax/neon-p5.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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/raddstoreexpminusmax.h>


void xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2(
    size_t elements,
    const float* input,
    const float* max,
    float* output,
    float* sum,
    const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(elements % sizeof(float) == 0);

  const float32x4_t vi_max = vld1q_dup_f32(max);
  const float32x4_t vlog2e = vld1q_dup_f32(&params->neonfma_rr1_p5.log2e);
  const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neonfma_rr1_p5.magic_bias);
  const float32x4_t vminus_ln2 = vld1q_dup_f32(&params->neonfma_rr1_p5.minus_ln2);
  const float32x4_t vc5 = vld1q_dup_f32(&params->neonfma_rr1_p5.c5);
  const float32x4_t vc4 = vld1q_dup_f32(&params->neonfma_rr1_p5.c4);
  const float32x4_t vc3 = vld1q_dup_f32(&params->neonfma_rr1_p5.c3);
  const float32x4_t vc2 = vld1q_dup_f32(&params->neonfma_rr1_p5.c2);
  const float32x4_t vc1 = vld1q_dup_f32(&params->neonfma_rr1_p5.c1);
  const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neonfma_rr1_p5.denorm_cutoff);

  float32x4_t vacc0 = vmovq_n_f32(0.0f);
  float32x4_t vacc1 = vmovq_n_f32(0.0f);
  for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
    const float32x4_t vi0123 = vld1q_f32(input); input += 4;
    const float32x4_t vi4567 = vld1q_f32(input); input += 4;

    const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max);
    const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max);

    float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e);
    float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e);

    const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
    const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));

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

    float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2);
    float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2);

    float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
    float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);

    vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
    vp4567 = vfmaq_f32(vc3, vp4567, vt4567);

    vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
    vp4567 = vfmaq_f32(vc2, vp4567, vt4567);

    vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
    vp4567 = vfmaq_f32(vc1, vp4567, vt4567);

    vt0123 = vmulq_f32(vt0123, vs0123);
    vt4567 = vmulq_f32(vt4567, vs4567);

    float32x4_t vf0123 = vfmaq_f32(vs0123, vp0123, vt0123);
    float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567);

    vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_cutoff)));
    vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_cutoff)));

    vst1q_f32(output, vf0123); output += 4;
    vst1q_f32(output, vf4567); output += 4;

    vacc0 = vaddq_f32(vacc0, vf0123);
    vacc0 = vaddq_f32(vacc0, vf4567);
  }
  vacc0 = vaddq_f32(vacc0, vacc1);

  float32x4_t vacc = vacc0;
  for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
    const float32x4_t vi = vld1q_f32(input); input += 4;

    const float32x4_t vx = vsubq_f32(vi, vi_max);

    float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e);

    const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

    vn = vsubq_f32(vn, vmagic_bias);

    float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2);

    float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
    vp = vfmaq_f32(vc3, vp, vt);
    vp = vfmaq_f32(vc2, vp, vt);
    vp = vfmaq_f32(vc1, vp, vt);

    vt = vmulq_f32(vt, vs);
    float32x4_t vf = vfmaq_f32(vs, vp, vt);

    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

    vst1q_f32(output, vf); output += 4;

    vacc = vaddq_f32(vacc, vf);
  }
#if XNN_ARCH_ARM64
  float vacc_lo = vaddvq_f32(vacc);
#else
  float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
#endif
  if (elements != 0) {
    assert(elements >= 1 * sizeof(float));
    assert(elements <= 3 * sizeof(float));
    const float32x4_t vi = vld1q_f32(input); input += 4;

    const float32x4_t vx = vsubq_f32(vi, vi_max);

    float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e);

    const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

    vn = vsubq_f32(vn, vmagic_bias);

    float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2);

    float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
    vp = vfmaq_f32(vc3, vp, vt);
    vp = vfmaq_f32(vc2, vp, vt);
    vp = vfmaq_f32(vc1, vp, vt);

    vt = vmulq_f32(vt, vs);
    float32x4_t vf = vfmaq_f32(vs, vp, vt);

    vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

    float32x2_t vf_lo = vget_low_f32(vf);
    if (elements & (2 * sizeof(float))) {
      vst1_f32(output, vf_lo); output += 2;

      #if XNN_ARCH_ARM64
        vacc_lo += vaddv_f32(vf_lo);
      #else
        vacc_lo = vadd_f32(vacc_lo, vf_lo);
      #endif

      vf_lo = vget_high_f32(vf);
    }
    if (elements & (1 * sizeof(float))) {
      vst1_lane_f32(output, vf_lo, 0);

      #if XNN_ARCH_ARM64
        vacc_lo += vget_lane_f32(vf_lo, 0);
      #else
        vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
      #endif
    }
  }
#if XNN_ARCH_ARM64
  *sum = vacc_lo;
#else
  vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
#endif
}