// 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 % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"]
#include <assert.h>

#include <immintrin.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__avx_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
    const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(input);
    $for N in range(8, BATCH_TILE, 8):
      const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(input + ${N});
    input += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      const __m256 vy${ABC[N:N+8]} = _mm256_round_ps(vx${ABC[N:N+8]}, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);

    _mm256_storeu_ps(output, vy${ABC[0:8]});
    $for N in range(8, BATCH_TILE, 8):
      _mm256_storeu_ps(output + ${N}, vy${ABC[N:N+8]});
    output += ${BATCH_TILE};
  }
  $if BATCH_TILE > 8:
    for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
      const __m256 vx = _mm256_loadu_ps(input);
      input += 8;

      const __m256 vy = _mm256_round_ps(vx, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);

      _mm256_storeu_ps(output, vy);
      output += 8;
    }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));

    const __m256 vx = _mm256_maskload_ps(input, vmask);
    const __m256 vy = _mm256_round_ps(vx, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC);

    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (batch & (4 * sizeof(float))) {
      _mm_storeu_ps(output, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      output += 4;
    }
    if (batch & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) output, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      _mm_store_ss(output, vy_lo);
    }
  }
}