// Auto-generated file. Do not edit!
//   Template: src/f16-vtanh/avx-expm1minus.c.in
//   Generator: tools/xngen
//
// Copyright 2023 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 <stdint.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/microparams.h>
#include <xnnpack/vunary.h>

void xnn_f16_vtanh_ukernel__avx2_expm1minus_rr1_p3h2ts_rcp_x8(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_tanh_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->avx_expm1minus_rr1_p3h2.sign_mask);
  const __m256 vsat_cutoff = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.sat_cutoff);
  const __m256 vlog2e = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.log2e);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.magic_bias);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.minus_ln2);
  const __m256 vc3 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.c3);
  const __m256 vc2 = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.c2);
  const __m256 vtwo = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.two);
  const __m256 vminus_one = _mm256_load_ps(params->avx_expm1minus_rr1_p3h2.minus_one);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) i);
    i += 8;

    const __m128i vabsx = _mm_or_si128(vx, vsign_mask);
    __m256 vz = _mm256_cvtph_ps(vabsx);

    const __m128i vinvsignx = _mm_xor_si128(vx, vabsx);
    const __m256 vm = _mm256_cmp_ps(vz, vsat_cutoff, _CMP_LE_OS);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = vc3;
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vtwo);

    const __m256 vts = _mm256_mul_ps(vt, vs);
    const __m256 vsmo = _mm256_add_ps(vs, vminus_one);
    const __m256 vemo = _mm256_fmadd_ps(vp, vts, vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vrepo = _mm256_rcp_ps(vepo);

    __m256 vy = _mm256_mul_ps(vemo, vrepo);

    vy = _mm256_blendv_ps(vy, vminus_one, vm);

    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT);
    vh = _mm_xor_si128(vh, vinvsignx);

    _mm_storeu_si128((__m128i*) o, vh);
    o += 8;
  }
  if (batch != 0) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) i);

    const __m128i vabsx = _mm_or_si128(vx, vsign_mask);
    __m256 vz = _mm256_cvtph_ps(vabsx);

    const __m128i vinvsignx = _mm_xor_si128(vx, vabsx);
    const __m256 vm = _mm256_cmp_ps(vz, vsat_cutoff, _CMP_LE_OS);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    const __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = vc3;
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vtwo);

    const __m256 vts = _mm256_mul_ps(vt, vs);
    const __m256 vsmo = _mm256_add_ps(vs, vminus_one);
    const __m256 vemo = _mm256_fmadd_ps(vp, vts, vsmo);

    const __m256 vepo = _mm256_add_ps(vemo, vtwo);

    __m256 vrepo = _mm256_rcp_ps(vepo);

    __m256 vy = _mm256_mul_ps(vemo, vrepo);

    vy = _mm256_blendv_ps(vy, vminus_one, vm);

    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT);
    vh = _mm_xor_si128(vh, vinvsignx);

    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}