// 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"
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_vhswish_ukernel__neonfp16arith_x${BATCH_TILE}(
    size_t batch,
    const void* restrict input,
    void* restrict output,
    const union xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;

  const float16x8_t vsixth = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.sixth));
  const float16x8_t vthree = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.three));
  const int16x8_t vsix = vreinterpretq_s16_u16(vld1q_dup_u16(&params->fp16arith.six));
  const int16x8_t vzero = vdupq_n_s16(0);

  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      $for N in range(0, BATCH_TILE, 8):
        float16x8_t vx${ABC[N:N+8]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

      $for N in range(0, BATCH_TILE, 8):
        float16x8_t vacc${ABC[N:N+8]} = vaddq_f16(vx${ABC[N:N+8]}, vthree);
        vx${ABC[N:N+8]} = vmulq_f16(vx${ABC[N:N+8]}, vsixth);

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vzero));

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc${ABC[N:N+8]}), vsix));

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = vmulq_f16(vacc${ABC[N:N+8]}, vx${ABC[N:N+8]});

      $for N in range(0, BATCH_TILE, 8):
        vst1q_u16(o, vreinterpretq_u16_f16(vacc${ABC[N:N+8]})); o += 8;
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);
    vst1q_u16(o, vreinterpretq_u16_f16(vacc)); o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i));
    float16x8_t vacc = vaddq_f16(vx, vthree);
    vx = vmulq_f16(vx, vsixth);
    vacc = vreinterpretq_f16_s16(vmaxq_s16(vreinterpretq_s16_f16(vacc), vzero));
    vacc = vreinterpretq_f16_s16(vminq_s16(vreinterpretq_s16_f16(vacc), vsix));
    vacc = vmulq_f16(vacc, vx);

    float16x4_t vacc_lo = vget_low_f16(vacc);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vacc_lo)); o += 4;
      vacc_lo = vget_high_f16(vacc);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
      vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_u16(o, vreinterpret_u16_f16(vacc_lo), 0);
    }
  }
}