// Copyright 2022 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 <wasm_simd128.h>

#include "xnnpack/common.h"
#include "xnnpack/lut.h"


void xnn_x8_lut_ukernel__wasmpshufb_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 v128_t vt0 = wasm_v128_load(table);
  $for T in range(1, 16):
    const v128_t vt${ABC[T]} = wasm_v128_load(table + ${T * 16});

  const v128_t vtable0 = vt0;
  $for T in range(1, 8):
    const v128_t vtable${ABC[T]} = wasm_v128_xor(vt${ABC[T-1]}, vt${ABC[T]});
  $for T in range(8, 16):
    const v128_t vtable${ABC[T]} = wasm_v128_xor(wasm_v128_xor(vt${ABC[T-1]}, vt${ABC[T]}), vtable${ABC[T-8]});

  const v128_t voffset = wasm_i8x16_const_splat(16);
  $if BATCH_TILE > 16:
    for (; batch >= ${BATCH_TILE} * sizeof(uint8_t); batch -= ${BATCH_TILE} * sizeof(uint8_t)) {
      v128_t vx0 = wasm_v128_load((const v128_t*) input);
      $for N in range(1, SIMD_TILE):
        v128_t vx${N} = wasm_v128_load((const v128_t*) (input + ${N * 16}));
      input += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        v128_t vy${N} = wasm_i8x16_relaxed_swizzle(vtable0, vx${N});

      $for T in range(1, 9):
        $for N in range(SIMD_TILE):
          vx${N} = wasm_i8x16_sub(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = wasm_v128_xor(vy${N}, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx${N}));

      $for T in range(9, 16):
        $for N in range(SIMD_TILE):
          vx${N} = wasm_i8x16_sub_sat(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = wasm_v128_xor(vy${N}, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx${N}));

      wasm_v128_store(output, vy0);
      $for N in range(1, SIMD_TILE):
        wasm_v128_store(output + ${N * 16}, vy${N});
      output += ${BATCH_TILE};
    }
  for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
    v128_t vx = wasm_v128_load(input);
    input += 16;

    v128_t vy = wasm_i8x16_relaxed_swizzle(vtable0, vx);

    $for T in range(1, 9):
      vx = wasm_i8x16_sub(vx, voffset);
      vy = wasm_v128_xor(vy, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = wasm_i8x16_sub_sat(vx, voffset);
      vy = wasm_v128_xor(vy, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx));

    wasm_v128_store(output, vy);
    output += 16;
  }
  if XNN_UNLIKELY(batch != 0) {
    v128_t vx = wasm_v128_load((const v128_t*) input);

    v128_t vy = wasm_i8x16_relaxed_swizzle(vtable0, vx);

    $for T in range(1, 9):
      vx = wasm_i8x16_sub(vx, voffset);
      vy = wasm_v128_xor(vy, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = wasm_i8x16_sub_sat(vx, voffset);
      vy = wasm_v128_xor(vy, wasm_i8x16_relaxed_swizzle(vtable${ABC[T]}, vx));

    if (batch & (8 * sizeof(uint8_t))) {
      wasm_v128_store64_lane(output, vy, 0);
      vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
      output += 8;
    }
    if (batch & (4 * sizeof(uint8_t))) {
      wasm_v128_store32_lane(output, vy, 0);
      vy = wasm_u64x2_shr(vy, 32);
      output += 4;
    }
    if (batch & (2 * sizeof(uint8_t))) {
      wasm_v128_store16_lane(output, vy, 0);
      vy = wasm_u32x4_shr(vy, 16);
      output += 2;
    }
    if (batch & (1 * sizeof(uint8_t))) {
      wasm_v128_store8_lane(output, vy, 0);
    }
  }
}