// Copyright 2019 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 <psimd.h>

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


void xnn_f32_hswish_ukernel__psimd_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_hswish_params params[restrict static 1])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const psimd_f32 vsixth = psimd_load_splat_f32(&params->scalar.sixth);
  const psimd_f32 vhalf = psimd_load_splat_f32(&params->scalar.half);
  const psimd_f32 vone = psimd_load_splat_f32(&params->scalar.one);
  const psimd_f32 vzero = psimd_splat_f32(0.0f);

  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    const psimd_f32 vx${ABC[0:4]} = psimd_load_f32(x);
    $for N in range(4, BATCH_TILE, 4):
      const psimd_f32 vx${ABC[N:N+4]} = psimd_load_f32(x + ${N});
    x += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 4):
      psimd_f32 vacc${ABC[N:N+4]} = psimd_qfma_f32(vhalf, vx${ABC[N:N+4]}, vsixth);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = psimd_max_f32(vacc${ABC[N:N+4]}, vzero);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = psimd_min_f32(vacc${ABC[N:N+4]}, vone);

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = psimd_mul_f32(vacc${ABC[N:N+4]}, vx${ABC[N:N+4]});

    psimd_store_f32(y, vacc${ABC[0:4]});
    $for N in range(4, BATCH_TILE, 4):
      psimd_store_f32(y + ${N}, vacc${ABC[N:N+4]});
    y += ${BATCH_TILE};
  }
  $if BATCH_TILE >= 4:
    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
      const psimd_f32 vx0123 = psimd_load_f32(x);
      x += 4;
      psimd_f32 vacc0123 = psimd_qfma_f32(vhalf, vx0123, vsixth);
      vacc0123 = psimd_max_f32(vacc0123, vzero);
      vacc0123 = psimd_min_f32(vacc0123, vone);
      vacc0123 = psimd_mul_f32(vacc0123, vx0123);
      psimd_store_f32(y, vacc0123);
      y += 4;
    }
  if XNN_UNLIKELY(n != 0) {
    const psimd_f32 vx0123 = psimd_load_f32(x);
    psimd_f32 vacc0123 = psimd_qfma_f32(vhalf, vx0123, vsixth);
    vacc0123 = psimd_max_f32(vacc0123, vzero);
    vacc0123 = psimd_min_f32(vacc0123, vone);
    vacc0123 = psimd_mul_f32(vacc0123, vx0123);

    if (n & (2 * sizeof(float))) {
      psimd_store2_f32(y, vacc0123);
      vacc0123 = psimd_concat_hi_f32(vacc0123, vacc0123);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      psimd_store1_f32(y, vacc0123);
    }
  }
}