// Auto-generated file. Do not edit!
//   Template: src/f16-f32-vcvt/sse-int16.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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 <smmintrin.h>

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


void xnn_f16_f32_vcvt_ukernel__avx_int16_x32(
    size_t n,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int16.sign_mask);
  const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int16.exp_offset);
  const __m128 vexp_scale = _mm_load_ps(params->sse_int16.exp_scale);
  const __m128i vmagic_mask = _mm_load_si128((const __m128i*) params->sse_int16.magic_mask);
  const __m128 vmagic_bias = _mm_load_ps(params->sse_int16.magic_bias);
  const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int16.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  for (; n >= 32 * sizeof(uint16_t); n -= 32 * sizeof(uint16_t)) {
    const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
    const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8));
    const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16));
    const __m128i vh3 = _mm_loadu_si128((const __m128i*) (i + 24));
    i += 32;

    const __m128i vsign0 = _mm_and_si128(vh0, vsign_mask);
    const __m128i vsign1 = _mm_and_si128(vh1, vsign_mask);
    const __m128i vsign2 = _mm_and_si128(vh2, vsign_mask);
    const __m128i vsign3 = _mm_and_si128(vh3, vsign_mask);

    const __m128i vnonsign0 = _mm_xor_si128(vh0, vsign0);
    const __m128i vnonsign1 = _mm_xor_si128(vh1, vsign1);
    const __m128i vnonsign2 = _mm_xor_si128(vh2, vsign2);
    const __m128i vnonsign3 = _mm_xor_si128(vh3, vsign3);

    const __m128i vprenorm0 = _mm_slli_epi16(vnonsign0, 13);
    const __m128i vprenorm1 = _mm_add_epi16(_mm_srli_epi16(vnonsign0, 3), vexp_offset);
    const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13);
    const __m128i vprenorm3 = _mm_add_epi16(_mm_srli_epi16(vnonsign1, 3), vexp_offset);
    const __m128i vprenorm4 = _mm_slli_epi16(vnonsign2, 13);
    const __m128i vprenorm5 = _mm_add_epi16(_mm_srli_epi16(vnonsign2, 3), vexp_offset);
    const __m128i vprenorm6 = _mm_slli_epi16(vnonsign3, 13);
    const __m128i vprenorm7 = _mm_add_epi16(_mm_srli_epi16(vnonsign3, 3), vexp_offset);

    const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm0, vprenorm1)), vexp_scale));
    const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm0, vprenorm1)), vexp_scale));
    const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), vexp_scale));
    const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), vexp_scale));
    const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm4, vprenorm5)), vexp_scale));
    const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm4, vprenorm5)), vexp_scale));
    const __m128i vnorm6 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm6, vprenorm7)), vexp_scale));
    const __m128i vnorm7 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm6, vprenorm7)), vexp_scale));

    const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign0, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign1, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign2, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm6 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign3, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm7 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign3, vmagic_mask)), vmagic_bias));

    const __m128i vmask0 = _mm_cmpgt_epi16(vnonsign0, vdenorm_cutoff);
    const __m128i vmask1 = _mm_cmpgt_epi16(vnonsign1, vdenorm_cutoff);
    const __m128i vmask2 = _mm_cmpgt_epi16(vnonsign2, vdenorm_cutoff);
    const __m128i vmask3 = _mm_cmpgt_epi16(vnonsign3, vdenorm_cutoff);

    const __m128i vf0 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign0),
      _mm_blendv_epi8(vdenorm0, vnorm0, _mm_cvtepi16_epi32(vmask0)));
    const __m128i vf1 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign0),
      _mm_blendv_epi8(vdenorm1, vnorm1, _mm_unpackhi_epi16(vmask0, vmask0)));
    const __m128i vf2 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign1),
      _mm_blendv_epi8(vdenorm2, vnorm2, _mm_cvtepi16_epi32(vmask1)));
    const __m128i vf3 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign1),
      _mm_blendv_epi8(vdenorm3, vnorm3, _mm_unpackhi_epi16(vmask1, vmask1)));
    const __m128i vf4 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign2),
      _mm_blendv_epi8(vdenorm4, vnorm4, _mm_cvtepi16_epi32(vmask2)));
    const __m128i vf5 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign2),
      _mm_blendv_epi8(vdenorm5, vnorm5, _mm_unpackhi_epi16(vmask2, vmask2)));
    const __m128i vf6 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign3),
      _mm_blendv_epi8(vdenorm6, vnorm6, _mm_cvtepi16_epi32(vmask3)));
    const __m128i vf7 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign3),
      _mm_blendv_epi8(vdenorm7, vnorm7, _mm_unpackhi_epi16(vmask3, vmask3)));

    _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
    _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1));
    _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2));
    _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3));
    _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4));
    _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5));
    _mm_storeu_ps(output + 24, _mm_castsi128_ps(vf6));
    _mm_storeu_ps(output + 28, _mm_castsi128_ps(vf7));
    output += 32;
  }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);
    i += 8;

    const __m128i vsign = _mm_and_si128(vh, vsign_mask);

    const __m128i vnonsign = _mm_xor_si128(vh, vsign);

    const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
    const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));

    const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);

    const __m128i vf_lo = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));

    const __m128i vf_hi = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));

    _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
    _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
    output += 8;
  }
  if XNN_UNPREDICTABLE(n != 0) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);

    const __m128i vsign = _mm_and_si128(vh, vsign_mask);

    const __m128i vnonsign = _mm_xor_si128(vh, vsign);

    const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13);
    const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias));

    const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff);

    __m128i vf = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign),
      _mm_blendv_epi8(vdenorm_lo, vnorm_lo, _mm_cvtepi16_epi32(vmask)));

    if (n & (4 * sizeof(uint16_t))) {
      _mm_storeu_ps(output, _mm_castsi128_ps(vf));
      output += 4;

      vf = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign),
        _mm_blendv_epi8(vdenorm_hi, vnorm_hi, _mm_unpackhi_epi16(vmask, vmask)));
    }
    if (n & (2 * sizeof(uint16_t))) {
      _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
      output += 2;

      vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
    }
    if (n & (1 * sizeof(uint16_t))) {
      _mm_store_ss(output, _mm_castsi128_ps(vf));
    }
  }
}