// 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
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/raddstoreexpminusmax.h>


void xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x${BATCH_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t batch,
    const void* input,
    const void* max,
    void* output,
    void* sum,
    const union xnn_f16_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch % sizeof(uint16_t) == 0);

  const __m256 vi_max = _mm256_cvtph_ps(_mm_set1_epi16((short) *((const uint16_t*) max)));
  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p2.log2e);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p2.magic_bias);
  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 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;
  $for K in range(ACCUMULATORS):
    __m256 vacc${K} = _mm256_setzero_ps();
  for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
    const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    $for N in range(1, SIMD_TILE):
      const __m256 vi${ABC[N]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
    i += ${BATCH_TILE};

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

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

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

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

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _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};

    $for N in range(SIMD_TILE):
      vacc${N % ACCUMULATORS} = _mm256_add_ps(vacc${N % ACCUMULATORS}, vf${ABC[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 (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 vi = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 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);

    const __m256 vp = _mm256_fmadd_ps(vc2, 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);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC));
    o += 8;

    vacc = _mm256_add_ps(vacc, vf);
  }
  __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
  if (batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));

    const __m256 vi = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    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);

    const __m256 vp = _mm256_fmadd_ps(vc2, 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);

    __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC);
    __m128 vf_lo = _mm256_castps256_ps128(vf);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      vacc_lo = _mm_add_ps(vacc_lo, vf_lo);
      vf_lo = _mm256_extractf128_ps(vf, 1);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      vacc_lo = _mm_blend_ps(_mm_add_ps(vacc_lo, vf_lo), vacc_lo, 0xC);
      vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
      vacc_lo = _mm_add_ss(vacc_lo, vf_lo);
    }
  }
  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));
  *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0);
  _mm256_zeroupper();
}