// Copyright 2019 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 ELEMENTS_TILE % 8 == 0
$assert ELEMENTS_TILE >= 8
$SIMD_TILE = ELEMENTS_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/raddstoreexpminusmax.h>


void xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    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)])
{
  assert(elements % sizeof(float) == 0);

  const __m256 vi_max = _mm256_broadcast_ss(max);
  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
  const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
  const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
  const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
  const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
  const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
  const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);

  $for K in range(ACCUMULATORS):
    __m256 vacc${K} = _mm256_setzero_ps();
  for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
    const __m256 vi0 = _mm256_loadu_ps(input);
    $for N in range(1, SIMD_TILE):
      const __m256 vi${N} = _mm256_loadu_ps(input + ${N * 8});
    input += ${ELEMENTS_TILE};

    $for N in range(SIMD_TILE):
      const __m256 vx${N} = _mm256_sub_ps(vi${N}, vi_max);

    $for N in range(SIMD_TILE):
      __m256 vn${N} = _mm256_fmadd_ps(vx${N}, vlog2e, vmagic_bias);

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

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

    $for N in range(SIMD_TILE):
      __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vx${N});

    $for N in range(SIMD_TILE):
      __m256 vp${N} = _mm256_fmadd_ps(vc5, vt${N}, vc4);

    $for N in range(SIMD_TILE):
      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc3);

    $for N in range(SIMD_TILE):
      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2);

    $for N in range(SIMD_TILE):
      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc1);

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

    $for N in range(SIMD_TILE):
      __m256 vf${N} = _mm256_fmadd_ps(vt${N}, vp${N}, vs${N});

    $for N in range(SIMD_TILE):
      vf${N} = _mm256_andnot_ps(_mm256_cmp_ps(vx${N}, vdenorm_cutoff, _CMP_LT_OS), vf${N});

    _mm256_storeu_ps(output, vf0);
    $for N in range(1, SIMD_TILE):
      _mm256_storeu_ps(output + ${N * 8}, vf${N});
    output += ${ELEMENTS_TILE};

    $for N in range(SIMD_TILE):
      vacc${N % ACCUMULATORS} = _mm256_add_ps(vacc${N % ACCUMULATORS}, vf${N});
  }
  $if ACCUMULATORS > 1:
    $ACC_SLICE = 1
    $while ACC_SLICE < ACCUMULATORS:
      $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
        $if A + ACC_SLICE < ACCUMULATORS:
          vacc${A} = _mm256_add_ps(vacc${A}, vacc${A + ACC_SLICE});
      $ACC_SLICE *= 2

  __m256 vacc = vacc0;
  for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
    const __m256 vi = _mm256_loadu_ps(input);
    input += 8;

    const __m256 vx = _mm256_sub_ps(vi, vi_max);

    __m256 vn = _mm256_fmadd_ps(vx, 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, vx);

    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
    vp = _mm256_fmadd_ps(vp, vt, vc3);
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);

    vt = _mm256_mul_ps(vt, vs);
    __m256 vf = _mm256_fmadd_ps(vt, vp, vs);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);

    _mm256_storeu_ps(output, vf);
    output += 8;

    vacc = _mm256_add_ps(vacc, vf);
  }
  if (elements != 0) {
    assert(elements >= 1 * sizeof(float));
    assert(elements <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx2_rr1_p5.mask_table[7] - elements));

    const __m256 vi = _mm256_maskload_ps(input, vmask);

    const __m256 vx = _mm256_sub_ps(vi, vi_max);

    __m256 vn = _mm256_fmadd_ps(vx, 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, vx);

    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
    vp = _mm256_fmadd_ps(vp, vt, vc3);
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);

    vt = _mm256_mul_ps(vt, vs);
    __m256 vf = _mm256_fmadd_ps(vt, vp, vs);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);

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

    vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
  }
  __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
  vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
  vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
  _mm_store_ss(sum, vacc_lo);
  _mm256_zeroupper();
}