// 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
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>

#include <immintrin.h>

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


extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16];

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

void xnn_f32_velu_ukernel__avx_rr2_lut16_p3_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 % sizeof(float) == 0);

  const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
  const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
  const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);

  const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
  const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p19f);
  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
  const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0xF));
  const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
  const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
  const __m256 vc3 = _mm256_set1_ps(0x1.55561Cp-3f);
  const __m256 vc2 = _mm256_set1_ps(0x1.0001ECp-1f);
  const __m256 vone = _mm256_set1_ps(1.0f);

  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      __m256 vx0 = _mm256_loadu_ps(x);
      $for N in range(1, SIMD_TILE):
        __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));

      $for N in range(SIMD_TILE):
        __m256 vn${N} = _mm256_add_ps(_mm256_mul_ps(vz${N}, vlog2e), vmagic_bias);

      $for N in range(SIMD_TILE):
        const __m256 vidx${N} = _mm256_and_ps(vn${N}, vindex_mask);

        const __m128i vidx${N}_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx${N})), 2);
        const __m128i vidx${N}_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx${N}, 1)), 2);
        #if XNN_ARCH_X86_64
          const uint64_t vidx${N}_ll = (uint64_t) _mm_cvtsi128_si64(vidx${N}_lo);
          const uint64_t vidx${N}_lh = (uint64_t) _mm_extract_epi64(vidx${N}_lo, 1);
          const uint64_t vidx${N}_hl = (uint64_t) _mm_cvtsi128_si64(vidx${N}_hi);
          const uint64_t vidx${N}_hh = (uint64_t) _mm_extract_epi64(vidx${N}_hi, 1);
          __m128i vl${N}_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${N}_ll));
          __m128i vl${N}_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${N}_lh));
          __m128i vl${N}_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${N}_hl));
          __m128i vl${N}_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${N}_hh));
          vl${N}_ll = _mm_insert_epi32(vl${N}_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${N}_ll >> 32))), 1);
          vl${N}_lh = _mm_insert_epi32(vl${N}_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${N}_lh >> 32))), 1);
          vl${N}_hl = _mm_insert_epi32(vl${N}_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${N}_hl >> 32))), 1);
          vl${N}_hh = _mm_insert_epi32(vl${N}_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${N}_hh >> 32))), 1);
        #else
          __m128i vl${N}_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx${N}_lo)));
          __m128i vl${N}_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo, 2)));
          __m128i vl${N}_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx${N}_hi)));
          __m128i vl${N}_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi, 2)));
          vl${N}_ll = _mm_insert_epi32(vl${N}_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo, 1))), 1);
          vl${N}_lh = _mm_insert_epi32(vl${N}_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo, 3))), 1);
          vl${N}_hl = _mm_insert_epi32(vl${N}_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi, 1))), 1);
          vl${N}_hh = _mm_insert_epi32(vl${N}_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi, 3))), 1);
        #endif
        const __m128i vl${N}_lo = _mm_unpacklo_epi64(vl${N}_ll, vl${N}_lh);
        const __m128i vl${N}_hi = _mm_unpacklo_epi64(vl${N}_hl, vl${N}_hh);

      $for N in range(SIMD_TILE):
        const __m128i ven${N}_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn${N})), 19);
        const __m128i ven${N}_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn${N}, 1)), 19);
        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
        const __m128 vs${N}_lo = _mm_castsi128_ps(_mm_add_epi32(vl${N}_lo, ven${N}_lo));
        const __m128 vs${N}_hi = _mm_castsi128_ps(_mm_add_epi32(vl${N}_hi, ven${N}_hi));

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

      $for N in range(SIMD_TILE):
        vt${N} = _mm256_add_ps(_mm256_mul_ps(vn${N}, vminus_ln2_lo), vt${N});
        __m256 vs${N} = _mm256_insertf128_ps(_mm256_castps128_ps256(vs${N}_lo), vs${N}_hi, 1);

      $for N in range(SIMD_TILE):
        __m256 vp${N} = _mm256_add_ps(_mm256_mul_ps(vc3, vt${N}), vc2);

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

      $for N in range(SIMD_TILE):
        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
        vs${N} = _mm256_sub_ps(vs${N}, vone);

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

      $for N in range(SIMD_TILE):
        const __m256 ve${N} = _mm256_mul_ps(_mm256_add_ps(vp${N}, vs${N}), valpha);
        vx${N} = _mm256_mul_ps(vx${N}, vbeta);

      $for N in range(SIMD_TILE):
        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});

      _mm256_storeu_ps(y, vy0);
      $for N in range(1, SIMD_TILE):
        _mm256_storeu_ps(y + ${N * 8}, vy${N});
      y += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    __m256 vx = _mm256_loadu_ps(x);
    x += 8;

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

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

    const __m256 vidx = _mm256_and_ps(vn, vindex_mask);

    const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
    const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
    #if XNN_ARCH_X86_64
      const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
      const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
      const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
      const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
      __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_ll));
      __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lh));
      __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hl));
      __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh));
      vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_ll >> 32))), 1);
      vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lh >> 32))), 1);
      vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hl >> 32))), 1);
      vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
    #else
      __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo)));
      __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2)));
      __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi)));
      __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2)));
      vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
      vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
      vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
      vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
    #endif
    const __m128i ven_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 19);
    const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19);

    const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
    const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);

    vn = _mm256_sub_ps(vn, vmagic_bias);
    const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ven_lo));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ven_hi));

    __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 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

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

    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_sub_ps(vs, vone);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);

    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

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

    __m256 vx = _mm256_maskload_ps(x, vmask);

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

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

    const __m256 vidx = _mm256_and_ps(vn, vindex_mask);

    const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
    const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
    #if XNN_ARCH_X86_64
      const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
      const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
      const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
      const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
      __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_ll));
      __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lh));
      __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hl));
      __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh));
      vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_ll >> 32))), 1);
      vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lh >> 32))), 1);
      vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hl >> 32))), 1);
      vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
    #else
      __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo)));
      __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2)));
      __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi)));
      __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2)));
      vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
      vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
      vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
      vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
    #endif
    const __m128i ven_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 19);
    const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19);

    const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
    const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);

    vn = _mm256_sub_ps(vn, vmagic_bias);
    const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ven_lo));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ven_hi));

    __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 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);

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

    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_sub_ps(vs, vone);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);

    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (n & (4 * sizeof(float))) {
      _mm_storeu_ps(y, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      y += 4;
    }
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy_lo);
    }
  }
}