// Copyright 2021 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 >= 64
$assert BATCH_TILE % 64 == 0
$SIMD_TILE = BATCH_TILE // 64
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


void xnn_x8_lut_ukernel__avx512skx_vpshufb_x${BATCH_TILE}(
    size_t n,
    const uint8_t* x,
    uint8_t* y,
    const uint8_t t[restrict XNN_MIN_ELEMENTS(256)])
{
  assert(n != 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m512i vt0 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) t));
  $for T in range(1, 16):
    const __m512i vt${ABC[T]} = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + ${T * 16})));

  const __m512i vtable0 = vt0;
  $for T in range(1, 8):
    const __m512i vtable${ABC[T]} = _mm512_xor_si512(vt${ABC[T-1]}, vt${ABC[T]});
  $for T in range(8, 16):
    const __m512i vtable${ABC[T]} = _mm512_xor_si512(_mm512_xor_si512(vt${ABC[T-1]}, vt${ABC[T]}), vtable${ABC[T-8]});

  const __m512i voffset = _mm512_set1_epi8(16);
  $if BATCH_TILE > 64:
    for (; n >= ${BATCH_TILE} * sizeof(uint8_t); n -= ${BATCH_TILE} * sizeof(uint8_t)) {
      __m512i vx0 = _mm512_loadu_si512(x);
      $for N in range(1, SIMD_TILE):
        __m512i vx${N} = _mm512_loadu_si512(x + ${N * 64});
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        __m512i vy${N} = _mm512_shuffle_epi8(vtable0, vx${N});

      $for T in range(1, 9):
        $for N in range(SIMD_TILE):
          vx${N} = _mm512_sub_epi8(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = _mm512_xor_si512(vy${N}, _mm512_shuffle_epi8(vtable${ABC[T]}, vx${N}));

      $for T in range(9, 16):
        $for N in range(SIMD_TILE):
          vx${N} = _mm512_subs_epi8(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = _mm512_xor_si512(vy${N}, _mm512_shuffle_epi8(vtable${ABC[T]}, vx${N}));

      _mm512_storeu_si512(y, vy0);
      $for N in range(1, SIMD_TILE):
        _mm512_storeu_si512(y + ${N * 64}, vy${N});
      y += ${BATCH_TILE};
    }
  for (; n >= 64 * sizeof(uint8_t); n -= 64 * sizeof(uint8_t)) {
    __m512i vx = _mm512_loadu_si512(x);
    x += 64;

    __m512i vy = _mm512_shuffle_epi8(vtable0, vx);

    $for T in range(1, 9):
      vx = _mm512_sub_epi8(vx, voffset);
      vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = _mm512_subs_epi8(vx, voffset);
      vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable${ABC[T]}, vx));

    _mm512_storeu_si512(y, vy);
    y += 64;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n < 64);
    const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << n) - UINT64_C(1)));

    __m512i vx = _mm512_maskz_loadu_epi8(vmask, x);

    __m512i vy = _mm512_shuffle_epi8(vtable0, vx);

    $for T in range(1, 9):
      vx = _mm512_sub_epi8(vx, voffset);
      vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = _mm512_subs_epi8(vx, voffset);
      vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable${ABC[T]}, vx));

    _mm512_mask_storeu_epi8(y, vmask, vy);
  }
}