// 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_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

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

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

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