// 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.

$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$assert DIV_ALGO in ["div", "nr2"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>

#include <immintrin.h>

#include "xnnpack/common.h"
#include "xnnpack/vunary.h"


void xnn_f32_vsigmoid_ukernel__avx_rr2_p5_${DIV_ALGO}_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
  const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
  const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
  const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
  const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
  const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
  const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
  const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
  const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
  const __m256 vone = _mm256_set1_ps(1.0f);
  const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);

  XNN_FORCE_REALIZATION(vsign_mask);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vminus_ln2_hi);
  XNN_FORCE_REALIZATION(vminus_ln2_lo);
  XNN_FORCE_REALIZATION(vc5);
  XNN_FORCE_REALIZATION(vc4);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vc1);
  XNN_FORCE_REALIZATION(vone);
  XNN_FORCE_REALIZATION(vdenorm_cutoff);

  $if DIV_ALGO == "nr2":
    const __m256 vtwo = _mm256_set1_ps(2.0f);
    XNN_FORCE_REALIZATION(vtwo);

  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const __m256 vx${ABC[0]} = _mm256_loadu_ps(input);
      $for N in range(1, SIMD_TILE):
        const __m256 vx${ABC[N]} = _mm256_loadu_ps(input + ${N * 8});
      input += ${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_add_ps(_mm256_mul_ps(vz${ABC[N]}, vlog2e), vmagic_bias);

      $for N in range(SIMD_TILE):
        const __m128 vs_lo${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn${ABC[N]})), 23));
        const __m128 vs_hi${ABC[N]} = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn${ABC[N]}, 1)), 23));
        const __m256 vs${ABC[N]} = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo${ABC[N]}), vs_hi${ABC[N]}, 1);

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

      $for N in range(SIMD_TILE):
        __m256 vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_hi), vz${ABC[N]});

      $for N in range(SIMD_TILE):
        vt${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vn${ABC[N]}, vminus_ln2_lo), vt${ABC[N]});

      $for N in range(SIMD_TILE):
        __m256 vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vc5, vt${ABC[N]}), vc4);

      $for N in range(SIMD_TILE):
        vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc3);

      $for N in range(SIMD_TILE):
        vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${ABC[N]}), vc2);

      $for N in range(SIMD_TILE):
        vp${ABC[N]} = _mm256_add_ps(_mm256_mul_ps(vp${ABC[N]}, vt${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_add_ps(_mm256_mul_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:
        $for N in range(SIMD_TILE):
          __m256 vr${ABC[N]} = _mm256_rcp_ps(vd${ABC[N]});

        $for N in range(SIMD_TILE):
          vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, vd${ABC[N]})));
          vr${ABC[N]} = _mm256_mul_ps(vr${ABC[N]}, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr${ABC[N]}, 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]});

      _mm256_storeu_ps(output, vf${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm256_storeu_ps(output + ${N * 8}, vf${ABC[N]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const __m256 vx = _mm256_loadu_ps(input);
    input += 8;

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);

    const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
    const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    $if DIV_ALGO == "div":
      __m256 vf = _mm256_div_ps(ve, vd);
    $else:
      __m256 vr = _mm256_rcp_ps(vd);
      vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
      vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, 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);

    _mm256_storeu_ps(output, vf);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));

    const __m256 vx = _mm256_maskload_ps(input, vmask);

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
    const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
    const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);

    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    $if DIV_ALGO == "div":
      __m256 vf = _mm256_div_ps(ve, vd);
    $else:
      __m256 vr = _mm256_rcp_ps(vd);
      vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
      vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, 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);

    __m128 vf_lo = _mm256_castps256_ps128(vf);
    if (batch & (4 * sizeof(float))) {
      _mm_storeu_ps(output, vf_lo);
      vf_lo = _mm256_extractf128_ps(vf, 1);
      output += 4;
    }
    if (batch & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) output, vf_lo);
      vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      _mm_store_ss(output, vf_lo);
    }
  }
}