// 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 >= 32
$assert BATCH_TILE % 32 == 0
$SIMD_TILE = BATCH_TILE // 32
$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__avx2_u${BATCH_TILE}(
    size_t batch,
    const uint8_t* input,
    uint8_t* output,
    const uint8_t table[restrict XNN_MIN_ELEMENTS(256)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint8_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m256i vt0 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) table));
  $for T in range(1, 16):
    const __m256i vt${ABC[T]} = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (table + ${T * 16})));

  const __m256i vtable0 = vt0;
  $for T in range(1, 8):
    const __m256i vtable${ABC[T]} = _mm256_xor_si256(vt${ABC[T-1]}, vt${ABC[T]});
  $for T in range(8, 16):
    const __m256i vtable${ABC[T]} = _mm256_xor_si256(_mm256_xor_si256(vt${ABC[T-1]}, vt${ABC[T]}), vtable${ABC[T-8]});

  const __m256i voffset = _mm256_set1_epi8(16);
  for (; batch >= ${BATCH_TILE} * sizeof(uint8_t); batch -= ${BATCH_TILE} * sizeof(uint8_t)) {
    __m256i vx0 = _mm256_loadu_si256((const __m256i*) input);
    $for N in range(1, SIMD_TILE):
      __m256i vx${N} = _mm256_loadu_si256((const __m256i*) (input + ${N * 32}));
    input += ${BATCH_TILE};

    $for N in range(SIMD_TILE):
      __m256i vy${N} = _mm256_shuffle_epi8(vtable0, vx${N});

    $for T in range(1, 9):
      $for N in range(SIMD_TILE):
        vx${N} = _mm256_sub_epi8(vx${N}, voffset);
      $for N in range(SIMD_TILE):
        vy${N} = _mm256_xor_si256(vy${N}, _mm256_shuffle_epi8(vtable${ABC[T]}, vx${N}));

    $for T in range(9, 16):
      $for N in range(SIMD_TILE):
        vx${N} = _mm256_subs_epi8(vx${N}, voffset);
      $for N in range(SIMD_TILE):
        vy${N} = _mm256_xor_si256(vy${N}, _mm256_shuffle_epi8(vtable${ABC[T]}, vx${N}));

    _mm256_storeu_si256((__m256i*) output, vy0);
    $for N in range(1, SIMD_TILE):
      _mm256_storeu_si256((__m256i*) (output + ${N * 32}), vy${N});
    output += ${BATCH_TILE};
  }
  for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
    __m128i vx = _mm_loadu_si128((const __m128i*) input);
    input += 16;

    __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);

    $for T in range(1, 9):
      vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable${ABC[T]}), vx));

    $for T in range(9, 16):
      vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable${ABC[T]}), vx));

    _mm_storeu_si128((__m128i*) output, vy);
    output += 16;
  }
  if XNN_UNLIKELY(batch != 0) {
    __m128i vx = _mm_loadu_si128((const __m128i*) input);

    __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);

    $for T in range(1, 9):
      vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable${ABC[T]}), vx));

    $for T in range(9, 16):
      vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable${ABC[T]}), vx));

    if (batch & (8 * sizeof(uint8_t))) {
      _mm_storel_epi64((__m128i*) output, vy);
      vy = _mm_unpackhi_epi64(vy, vy);
      output += 8;
    }
    if (batch & (4 * sizeof(uint8_t))) {
      _mm_storeu_si32(output, vy);
      vy = _mm_srli_epi64(vy, 32);
      output += 4;
    }
    if (batch & (2 * sizeof(uint8_t))) {
      _mm_storeu_si16(output, vy);
      vy = _mm_srli_epi32(vy, 16);
      output += 2;
    }
    if (batch & (1 * sizeof(uint8_t))) {
      *output = (uint8_t) _mm_extract_epi8(vy, 0);
    }
  }
}