// 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 >= 16
$assert BATCH_TILE % 16 == 0
$SIMD_TILE = BATCH_TILE // 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/lut.h"
#include "xnnpack/common.h"


void xnn_x8_lut_ukernel__aarch64_neon_tbx128x4_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 uint8x16x4_t vtable0123 = vld1q_u8_x4(table);
  const uint8x16x4_t vtable4567 = vld1q_u8_x4(table + 64);
  const uint8x16x4_t vtable89AB = vld1q_u8_x4(table + 128);
  const uint8x16x4_t vtableCDEF = vld1q_u8_x4(table + 192);
  const uint8x16_t voffset = vmovq_n_u8(64);
  $if BATCH_TILE > 16:
    for (; batch >= ${BATCH_TILE} * sizeof(uint8_t); batch -= ${BATCH_TILE} * sizeof(uint8_t)) {
      $for N in range(SIMD_TILE):
        uint8x16_t vx${N} = vld1q_u8(input); input += 16;

      $for N in range(SIMD_TILE):
        uint8x16_t vy${N} = vqtbl4q_u8(vtable0123, vx${N});
        vx${N} = vsubq_u8(vx${N}, voffset);

      $for N in range(SIMD_TILE):
        vy${N} = vqtbx4q_u8(vy${N}, vtable4567, vx${N});
        vx${N} = vsubq_u8(vx${N}, voffset);

      $for N in range(SIMD_TILE):
        vy${N} = vqtbx4q_u8(vy${N}, vtable89AB, vx${N});
        vx${N} = vsubq_u8(vx${N}, voffset);

      $for N in range(SIMD_TILE):
        vy${N} = vqtbx4q_u8(vy${N}, vtableCDEF, vx${N});

      $for N in range(SIMD_TILE):
        vst1q_u8(output, vy${N}); output += 16;
    }
  for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
    uint8x16_t vx = vld1q_u8(input); input += 16;

    uint8x16_t vy = vqtbl4q_u8(vtable0123, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtable4567, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtable89AB, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtableCDEF, vx);

    vst1q_u8(output, vy); output += 16;
  }
  if XNN_UNLIKELY(batch != 0) {
    uint8x16_t vx = vld1q_u8(input);

    uint8x16_t vy = vqtbl4q_u8(vtable0123, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtable4567, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtable89AB, vx);

    vx = vsubq_u8(vx, voffset);
    vy = vqtbx4q_u8(vy, vtableCDEF, vx);

    uint8x8_t vy_lo = vget_low_u8(vy);
    if (batch & (8 * sizeof(uint8_t))) {
      vst1_u8(output, vy_lo); output += 8;
      vy_lo = vget_high_u8(vy);
    }
    if (batch & (4 * sizeof(uint8_t))) {
      vst1_lane_u32((void*) output, vreinterpret_u32_u8(vy_lo), 0); output += 4;
      vy_lo = vext_u8(vy_lo, vy_lo, 4);
    }
    if (batch & (2 * sizeof(uint8_t))) {
      vst1_lane_u16((void*) output, vreinterpret_u16_u8(vy_lo), 0); output += 2;
      vy_lo = vext_u8(vy_lo, vy_lo, 2);
    }
    if (batch & (1 * sizeof(uint8_t))) {
      vst1_lane_u8(output, vy_lo, 0);
    }
  }
}