// 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"
#include <assert.h>

#include <emmintrin.h>

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


void xnn_f32_vrndd_ukernel__sse2_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);

  const __m128i vmagic = _mm_load_si128((const __m128i*) params->sse2.sign_mask);
  const __m128 vone = _mm_load_ps(params->sse2.one);
  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};

    $for N in range(0, BATCH_TILE, 4):
      const __m128i vintx${ABC[N:N+4]} = _mm_cvttps_epi32(vx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      const __m128 vrndmask${ABC[N:N+4]} = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx${ABC[N:N+4]}, vmagic)));

    $for N in range(0, BATCH_TILE, 4):
      const __m128 vprerndx${ABC[N:N+4]} = _mm_cvtepi32_ps(vintx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      const __m128 vrndx${ABC[N:N+4]} = _mm_or_ps(_mm_and_ps(vx${ABC[N:N+4]}, vrndmask${ABC[N:N+4]}), _mm_andnot_ps(vrndmask${ABC[N:N+4]}, vprerndx${ABC[N:N+4]}));

    $for N in range(0, BATCH_TILE, 4):
      const __m128 vy${ABC[N:N+4]} = _mm_sub_ps(vrndx${ABC[N:N+4]}, _mm_and_ps(_mm_cmpgt_ps(vrndx${ABC[N:N+4]}, vx${ABC[N:N+4]}), vone));

    _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 __m128i vintx = _mm_cvttps_epi32(vx);
      const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic)));
      const __m128 vprerndx = _mm_cvtepi32_ps(vintx);
      const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx));
      const __m128 vy = _mm_sub_ps(vrndx, _mm_and_ps(_mm_cmpgt_ps(vrndx, vx), vone));

      _mm_storeu_ps(y, vy);
      y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    const __m128 vx = _mm_loadu_ps(x);
    const __m128i vintx = _mm_cvttps_epi32(vx);
    const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic)));
    const __m128 vprerndx = _mm_cvtepi32_ps(vintx);
    const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx));
    __m128 vy = _mm_sub_ps(vrndx, _mm_and_ps(_mm_cmpgt_ps(vrndx, vx), vone));
    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);
    }
  }
}