// 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 % 16 == 0
$assert BATCH_TILE >= 16
$SIMD_TILE = BATCH_TILE // 16
#include <assert.h>

#include <immintrin.h>

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


void xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const __m512 vprescale = _mm512_set1_ps(params->avx512_rr1_lut16_p3.prescale);
  const __m512 valpha = _mm512_set1_ps(params->avx512_rr1_lut16_p3.alpha);
  const __m512 vbeta = _mm512_set1_ps(params->avx512_rr1_lut16_p3.beta);
  const __m512 vsat_cutoff = _mm512_set1_ps(params->avx512_rr1_lut16_p3.sat_cutoff);
  const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias);
  const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e);
  const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2);
  const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3);
  const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2);
  const __m512i vtable = _mm512_load_si512(params->avx512_rr1_lut16_p3.table);

  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      __m512 vx0 = _mm512_loadu_ps(x);
      $for N in range(1, SIMD_TILE):
        __m512 vx${N} = _mm512_loadu_ps(x + ${N * 16});
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m512 vz${N} = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx${N}, vprescale));

      $for N in range(SIMD_TILE):
        __m512 vn${N} = _mm512_fmadd_ps(vz${N}, vlog2e, vmagic_bias);

      $for N in range(SIMD_TILE):
        const __m512i ven${N} = _mm512_slli_epi32(_mm512_castps_si512(vn${N}), 19);
        const __m512i vl${N} = _mm512_permutexvar_epi32(_mm512_castps_si512(vn${N}), vtable);

      $for N in range(SIMD_TILE):
        __m512 vs${N} = _mm512_castsi512_ps(_mm512_add_epi32(vl${N}, ven${N}));
        vn${N} = _mm512_sub_ps(vn${N}, vmagic_bias);

      $for N in range(SIMD_TILE):
        __m512 vt${N} = _mm512_fmadd_ps(vn${N}, vminus_ln2, vz${N});

      $for N in range(SIMD_TILE):
        __m512 vp${N} = _mm512_fmadd_ps(vc3, vt${N}, vc2);

      $for N in range(SIMD_TILE):
        vp${N} = _mm512_mul_ps(vp${N}, vt${N});
        vt${N} = _mm512_mul_ps(vt${N}, vs${N});

      $for N in range(SIMD_TILE):
        vs${N} = _mm512_fmsub_ps(vs${N}, valpha, valpha);

      $for N in range(SIMD_TILE):
        vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vt${N});

      const __m512 vzero = _mm512_setzero_ps();
      $for N in range(SIMD_TILE):
        __m512 vy${N} = _mm512_fmadd_ps(vp${N}, valpha, vs${N});
        const __mmask16 vsign${N} = _mm512_cmp_ps_mask(vx${N}, vzero, _CMP_NLT_US);

      $for N in range(SIMD_TILE):
        vy${N} = _mm512_mask_mul_ps(vy${N}, vsign${N}, vx${N}, vbeta);

      _mm512_storeu_ps(y, vy0);
      $for N in range(1, SIMD_TILE):
        _mm512_storeu_ps(y + ${N * 16}, vy${N});
      y += ${BATCH_TILE};
    }
  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
    __m512 vx = _mm512_loadu_ps(x);
    x += 16;

    const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
    const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);

    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
    const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
    const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
    __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
    vn = _mm512_sub_ps(vn, vmagic_bias);

    __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);

    __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
    vp = _mm512_mul_ps(vp, vt);

    vt = _mm512_mul_ps(vt, vs);
    vs = _mm512_fmsub_ps(vs, valpha, valpha);
    vp = _mm512_fmadd_ps(vp, vt, vt);
    __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);

    vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);

    _mm512_storeu_ps(y, vy);
    y += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 15 * sizeof(float));
    // Prepare mask for valid 32-bit elements (depends on n).
    n >>= 2 /* log2(sizeof(float)) */;
    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));

    __m512 vx = _mm512_maskz_loadu_ps(vmask, x);

    const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
    const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);

    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
    const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
    const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
    __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
    vn = _mm512_sub_ps(vn, vmagic_bias);

    __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);

    __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
    vp = _mm512_mul_ps(vp, vt);

    vt = _mm512_mul_ps(vt, vs);
    vs = _mm512_fmsub_ps(vs, valpha, valpha);
    vp = _mm512_fmadd_ps(vp, vt, vt);
    __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);

    vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);

    _mm512_mask_storeu_ps(y, vmask, vy);
  }
}