// 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 % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


$ISA = {0: "avx", 3: "fma3"}[FMA]
void xnn_f32_vhswish_ukernel__${ISA}_u${BATCH_TILE}(
    size_t batch,
    const float* input,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

  const __m256 vsixth = _mm256_set1_ps(0x1.555556p-3f);
  const __m256 vhalf = _mm256_set1_ps(0.5f);
  const __m256 vone = _mm256_set1_ps(1.0f);
  const __m256 vzero = _mm256_setzero_ps();

  XNN_FORCE_REALIZATION(vsixth);
  XNN_FORCE_REALIZATION(vhalf);
  XNN_FORCE_REALIZATION(vone);
  // XNN_FORCE_REALIZATION(vzero);

  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(input);
      $for N in range(8, BATCH_TILE, 8):
        const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(input + ${N});
      input += ${BATCH_TILE};

      $if FMA == 3:
        $for N in range(0, BATCH_TILE, 8):
          __m256 vacc${ABC[N:N+8]} = _mm256_fmadd_ps(vx${ABC[N:N+8]}, vsixth, vhalf);
      $else:
        $for N in range(0, BATCH_TILE, 8):
          __m256 vacc${ABC[N:N+8]} = _mm256_mul_ps(vx${ABC[N:N+8]}, vsixth);

        $for N in range(0, BATCH_TILE, 8):
          vacc${ABC[N:N+8]} = _mm256_add_ps(vacc${ABC[N:N+8]}, vhalf);

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = _mm256_max_ps(vacc${ABC[N:N+8]}, vzero);

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = _mm256_min_ps(vacc${ABC[N:N+8]}, vone);

      $for N in range(0, BATCH_TILE, 8):
        vacc${ABC[N:N+8]} = _mm256_mul_ps(vacc${ABC[N:N+8]}, vx${ABC[N:N+8]});

      _mm256_storeu_ps(output, vacc${ABC[0:8]});
      $for N in range(8, BATCH_TILE, 8):
        _mm256_storeu_ps(output + ${N}, vacc${ABC[N:N+8]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    const __m256 vx = _mm256_loadu_ps(input);
    input += 8;
    $if FMA == 3:
      __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
    $else:
      __m256 vacc = _mm256_mul_ps(vx, vsixth);
      vacc = _mm256_add_ps(vacc, vhalf);
    vacc = _mm256_max_ps(vacc, vzero);
    vacc = _mm256_min_ps(vacc, vone);
    vacc = _mm256_mul_ps(vacc, vx);
    _mm256_storeu_ps(output, vacc);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch));

    const __m256 vx = _mm256_maskload_ps(input, vmask);
    $if FMA == 3:
      __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
    $else:
      __m256 vacc = _mm256_mul_ps(vx, vsixth);
      vacc = _mm256_add_ps(vacc, vhalf);
    vacc = _mm256_max_ps(vacc, vzero);
    vacc = _mm256_min_ps(vacc, vone);
    vacc = _mm256_mul_ps(vacc, vx);

    __m128 vacc_lo = _mm256_castps256_ps128(vacc);
    if (batch & (4 * sizeof(float))) {
      _mm_storeu_ps(output, vacc_lo);
      vacc_lo = _mm256_extractf128_ps(vacc, 1);
      output += 4;
    }
    if (batch & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) output, vacc_lo);
      vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      _mm_store_ss(output, vacc_lo);
    }
  }
}