// Copyright 2022 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.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$assert DIV_ALGO in ["div", "rcp"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>

#include <immintrin.h>

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


void xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_${DIV_ALGO}_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch % sizeof(uint16_t) == 0);

  const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p2.sign_mask);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p2.magic_bias);
  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p2.log2e);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p2.minus_ln2);
  const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p2.c2);
  const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p2.c1);
  const __m256 vone = _mm256_load_ps(params->avx2_rr1_p2.one);
  const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p2.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m256 vx${ABC[0]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
      $for N in range(1, SIMD_TILE):
        const __m256 vx${ABC[N]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m256 vz${ABC[N]} = _mm256_or_ps(vx${ABC[N]}, vsign_mask);

      $for N in range(SIMD_TILE):
        __m256 vn${ABC[N]} = _mm256_fmadd_ps(vz${ABC[N]}, vlog2e, vmagic_bias);

      $for N in range(SIMD_TILE):
        const __m256 vs${ABC[N]} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn${ABC[N]}), 23));

      $for N in range(SIMD_TILE):
        vn${ABC[N]} = _mm256_sub_ps(vn${ABC[N]}, vmagic_bias);

      $for N in range(SIMD_TILE):
        __m256 vt${ABC[N]} = _mm256_fmadd_ps(vn${ABC[N]}, vminus_ln2, vz${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m256 vp${ABC[N]} = _mm256_fmadd_ps(vc2, vt${ABC[N]}, vc1);

      $for N in range(SIMD_TILE):
        vt${ABC[N]} = _mm256_mul_ps(vt${ABC[N]}, vs${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m256 ve${ABC[N]} = _mm256_fmadd_ps(vt${ABC[N]}, vp${ABC[N]}, vs${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m256 vd${ABC[N]} = _mm256_add_ps(ve${ABC[N]}, vone);

      $if DIV_ALGO == "div":
        $for N in range(SIMD_TILE):
          __m256 vf${ABC[N]} = _mm256_div_ps(ve${ABC[N]}, vd${ABC[N]});
      $else:
        $for N in range(SIMD_TILE):
          const __m256 vr${ABC[N]} = _mm256_rcp_ps(vd${ABC[N]});

        $for N in range(SIMD_TILE):
          __m256 vf${ABC[N]} = _mm256_mul_ps(ve${ABC[N]}, vr${ABC[N]});

      $for N in range(SIMD_TILE):
        vf${ABC[N]} = _mm256_andnot_ps(_mm256_cmp_ps(vz${ABC[N]}, vdenorm_cutoff, _CMP_LT_OS), vf${ABC[N]});

      $for N in range(SIMD_TILE):
        vf${ABC[N]} = _mm256_blendv_ps(_mm256_sub_ps(vone, vf${ABC[N]}), vf${ABC[N]}, vx${ABC[N]});

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf${ABC[0]}, _MM_FROUND_NO_EXC));
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (o + ${N * 8}), _mm256_cvtps_ph(vf${ABC[N]}, _MM_FROUND_NO_EXC));
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __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);

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

    const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    $if DIV_ALGO == "div":
      __m256 vf = _mm256_div_ps(ve, vd);
    $else:
      const __m256 vr = _mm256_rcp_ps(vd);
      __m256 vf = _mm256_mul_ps(ve, vr);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
    vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
    o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __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);

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

    const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    $if DIV_ALGO == "div":
      __m256 vf = _mm256_div_ps(ve, vd);
    $else:
      const __m256 vr = _mm256_rcp_ps(vd);
      __m256 vf = _mm256_mul_ps(ve, vr);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
    vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

    __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
    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);
    }
  }
}