// 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 ["RNDNE", "RNDZ", "RNDU", "RNDD"]
#include <assert.h>

#include <smmintrin.h>

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


$_MM_FROUND_TO_FLAG = {
$  "RNDNE": "_MM_FROUND_TO_NEAREST_INT",
$  "RNDZ": "_MM_FROUND_TO_ZERO",
$  "RNDU": "_MM_FROUND_TO_POS_INF",
$  "RNDD": "_MM_FROUND_TO_NEG_INF",
$}[OP]
void xnn_f32_v${OP.lower()}_ukernel__sse41_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    const __m128 vx${ABC[0:4]} = _mm_loadu_ps(x);
    $for N in range(4, BATCH_TILE, 4):
      const __m128 vx${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
    x += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 4):
      const __m128 vy${ABC[N:N+4]} = _mm_round_ps(vx${ABC[N:N+4]}, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);

    _mm_storeu_ps(y, vy${ABC[0:4]});
    $for N in range(4, BATCH_TILE, 4):
      _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]});
    y += ${BATCH_TILE};
  }
  $if BATCH_TILE > 4:
    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
      const __m128 vx = _mm_loadu_ps(x);
      x += 4;

      const __m128 vy = _mm_round_ps(vx, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);

      _mm_storeu_ps(y, vy);
      y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    const __m128 vx = _mm_loadu_ps(x);
    __m128 vy = _mm_round_ps(vx, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy);
      vy = _mm_movehl_ps(vy, vy);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy);
    }
  }
}