// 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 % 16 == 0
$assert BATCH_TILE >= 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/hswish.h>


void xnn_f32_hswish_ukernel__avx512f_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const __m512 vsixth = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.sixth));
  const __m512 vhalf = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.half));
  const __m512 vone = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.one));
  const __m512 vzero = _mm512_setzero_ps();

  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    const __m512 vx${ABC[0:16]} = _mm512_loadu_ps(x);
    $for N in range(16, BATCH_TILE, 16):
      const __m512 vx${ABC[N:N+16]} = _mm512_loadu_ps(x + ${N});
    x += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 16):
      __m512 vacc${ABC[N:N+16]} = _mm512_fmadd_ps(vx${ABC[N:N+16]}, vsixth, vhalf);

    $for N in range(0, BATCH_TILE, 16):
      vacc${ABC[N:N+16]} = _mm512_max_ps(vacc${ABC[N:N+16]}, vzero);

    $for N in range(0, BATCH_TILE, 16):
      vacc${ABC[N:N+16]} = _mm512_min_ps(vacc${ABC[N:N+16]}, vone);

    $for N in range(0, BATCH_TILE, 16):
      vacc${ABC[N:N+16]} = _mm512_mul_ps(vacc${ABC[N:N+16]}, vx${ABC[N:N+16]});

    _mm512_storeu_ps(y, vacc${ABC[0:16]});
    $for N in range(16, BATCH_TILE, 16):
      _mm512_storeu_ps(y + ${N}, vacc${ABC[N:N+16]});
    y += ${BATCH_TILE};
  }
  $if BATCH_TILE > 16:
    for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
      const __m512 vx = _mm512_loadu_ps(x);
      x += 16;
      __m512 vacc = _mm512_fmadd_ps(vx, vsixth, vhalf);
      vacc = _mm512_max_ps(vacc, vzero);
      vacc = _mm512_min_ps(vacc, vone);
      vacc = _mm512_mul_ps(vacc, vx);
      _mm512_storeu_ps(y, vacc);
      y += 16;
    }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 15 * sizeof(float));
    // Prepare mask for valid 32-bit elements (depends on n).
    n >>= 2 /* log2(sizeof(float)) */;
    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));

    const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
    __m512 vacc = _mm512_fmadd_ps(vx, vsixth, vhalf);
    vacc = _mm512_max_ps(vacc, vzero);
    vacc = _mm512_min_ps(vacc, vone);
    vacc = _mm512_mul_ps(vacc, vx);
    _mm512_mask_storeu_ps(y, vmask, vacc);
  }
}