// 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 <wasm_simd128.h>

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


void xnn_x8_lut_ukernel__wasmsimd_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 v128_t vtable0 = wasm_v128_load(t);
  $for T in range(1, 16):
    const v128_t vtable${T} = wasm_v128_load(t + ${T * 16});
  const v128_t voffset = wasm_i8x16_const_splat(16);
  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(uint8_t); n -= ${BATCH_TILE} * sizeof(uint8_t)) {
      v128_t vx0 = wasm_v128_load(x);
      $for N in range(1, SIMD_TILE):
        v128_t vx${N} = wasm_v128_load(x + ${N * 16});
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        v128_t vy${N} = wasm_i8x16_swizzle(vtable0, vx${N});

      $for T in range(1, 16):
        $for N in range(SIMD_TILE):
          vx${N} = wasm_i8x16_sub(vx${N}, voffset);
          vy${N} = wasm_v128_or(vy${N}, wasm_i8x16_swizzle(vtable${T}, vx${N}));

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

    v128_t vy = wasm_i8x16_swizzle(vtable0, vx);

    $for T in range(1, 16):
      vx = wasm_i8x16_sub(vx, voffset);
      vy = wasm_v128_or(vy, wasm_i8x16_swizzle(vtable${T}, vx));

    wasm_v128_store(y, vy);
    y += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    v128_t vx = wasm_v128_load(x);

    v128_t vy = wasm_i8x16_swizzle(vtable0, vx);

    $for T in range(1, 16):
      vx = wasm_i8x16_sub(vx, voffset);
      vy = wasm_v128_or(vy, wasm_i8x16_swizzle(vtable${T}, vx));

    if (n & (8 * sizeof(uint8_t))) {
      *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
      vy = wasm_v64x2_shuffle(vy, vy, 1, 1);
      y += 8;
    }
    if (n & (4 * sizeof(uint8_t))) {
      *((float*) y) = wasm_f32x4_extract_lane(vy, 0);
      vy = wasm_u64x2_shr(vy, 32);
      y += 4;
    }
    uint32_t vy_lo = wasm_i32x4_extract_lane(vy, 0);
    if (n & (2 * sizeof(uint8_t))) {
      *((uint16_t*) y) = (uint16_t) vy_lo;
      vy_lo >>= 16;
      y += 2;
    }
    if (n & (1 * sizeof(uint8_t))) {
      *y = (uint8_t) vy_lo;
    }
  }
}