// 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 <arm_neon.h>

#include "xnnpack/common.h"
#include "xnnpack/math.h"
#include "xnnpack/vunary.h"


void xnn_f32_vrndd_ukernel__neon_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);

  const float32x4_t vintegral_threshold = vreinterpretq_f32_u32(vmovq_n_u32(UINT32_C(0x4B000000)));
  const uint32x4_t vone = vreinterpretq_u32_f32(vmovq_n_f32(1.0f));
  for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
    $for N in range(0, BATCH_TILE, 4):
      const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(input); input += 4;

    $for N in range(0, BATCH_TILE, 4):
      const int32x4_t vintx${ABC[N:N+4]} = vcvtq_s32_f32(vx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      uint32x4_t vrndmask${ABC[N:N+4]} = vcaltq_f32(vx${ABC[N:N+4]}, vintegral_threshold);

    $for N in range(0, BATCH_TILE, 4):
      const float32x4_t vprerndx${ABC[N:N+4]} = vcvtq_f32_s32(vintx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      vrndmask${ABC[N:N+4]} = vbicq_u32(vrndmask${ABC[N:N+4]}, vmovq_n_u32(UINT32_C(0x80000000)));

    $for N in range(0, BATCH_TILE, 4):
      const float32x4_t vrndx${ABC[N:N+4]} = vbslq_f32(vrndmask${ABC[N:N+4]}, vprerndx${ABC[N:N+4]}, vx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      const uint32x4_t vadjmask${ABC[N:N+4]} = vcgtq_f32(vrndx${ABC[N:N+4]}, vx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      const float32x4_t vadjrndx${ABC[N:N+4]} = vreinterpretq_f32_u32(vandq_u32(vadjmask${ABC[N:N+4]}, vone));

    $for N in range(0, BATCH_TILE, 4):
      const float32x4_t vy${ABC[N:N+4]} = vsubq_f32(vrndx${ABC[N:N+4]}, vadjrndx${ABC[N:N+4]});

    $for N in range(0, BATCH_TILE, 4):
      vst1q_f32(output, vy${ABC[N:N+4]}); output += 4;
  }
  $if BATCH_TILE > 4:
    for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
      const float32x4_t vx = vld1q_f32(input); input += 4;
      const int32x4_t vintx = vcvtq_s32_f32(vx);
      uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
      const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
      vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
      const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
      const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
      const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
      const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
      vst1q_f32(output, vy); output += 4;
    }
  if XNN_UNLIKELY(batch != 0) {
    const float32x4_t vx = vld1q_f32(input);
    const int32x4_t vintx = vcvtq_s32_f32(vx);
    uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
    const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
    vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
    const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
    const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
    const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
    const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
    float32x2_t vy_lo = vget_low_f32(vy);
    if (batch & (2 * sizeof(float))) {
      vst1_f32(output, vy_lo); output += 2;
      vy_lo = vget_high_f32(vy);
    }
    if (batch & (1 * sizeof(float))) {
      vst1_lane_f32(output, vy_lo, 0);
    }
  }
}