// Copyright 2020 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 % 4 == 0
$assert BATCH_TILE >= 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ABS", "NEG", "SQR"]
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/vunary.h>


$WASM_F32X4_OP = {
$  "ABS": lambda x: "wasm_v128_and(%s, vnonsign_mask)" % x,
$  "NEG": lambda x: "wasm_v128_xor(%s, vsign_mask)" % x,
$  "SQR": lambda x: "wasm_f32x4_mul(%s, %s)" % (x, x),
$}[OP]
$PARAMS = {
$  "ABS": "xnn_f32_abs_params",
$  "NEG": "xnn_f32_neg_params",
$  "SQR": "xnn_f32_default_params",
$}[OP]
void xnn_f32_v${OP.lower()}_ukernel__wasmsimd_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  $if OP == "ABS":
    const v128_t vnonsign_mask = wasm_v128_load64_splat(&params->wasmsimd.nonsign_mask);
  $elif OP == "NEG":
    const v128_t vsign_mask = wasm_v128_load64_splat(&params->wasmsimd.sign_mask);
  $if BATCH_TILE > 4:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      const v128_t vx${ABC[0:4]} = wasm_v128_load(x);
      $for N in range(4, BATCH_TILE, 4):
        const v128_t vx${ABC[N:N+4]} = wasm_v128_load(x + ${N});
      x += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        const v128_t vy${ABC[N:N+4]} = ${WASM_F32X4_OP("vx" + ABC[N:N+4])};

      wasm_v128_store(y, vy${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        wasm_v128_store(y + ${N}, vy${ABC[N:N+4]});
      y += ${BATCH_TILE};
    }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    const v128_t vx = wasm_v128_load(x);
    x += 4;

    const v128_t vy = ${WASM_F32X4_OP("vx")};

    wasm_v128_store(y, vy);
    y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    const v128_t vx = wasm_v128_load(x);
    v128_t vy = ${WASM_F32X4_OP("vx")};
    if (n & (2 * sizeof(float))) {
      *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
      vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      *y = wasm_f32x4_extract_lane(vy, 0);
    }
  }
}