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

#include <immintrin.h>

#include "xnnpack/common.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/vunary.h"


void xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_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);

  const __m512 vhalf = _mm512_set1_ps(0.5f);
  $if BATCH_TILE > 16:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const __m512 vx${ABC[0]} = _mm512_loadu_ps(input);
      $for N in range(1, SIMD_TILE):
        const __m512 vx${ABC[N]} = _mm512_loadu_ps(input + ${N * 16});
      input += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m512 vrsqrtx${ABC[N]} = _mm512_rsqrt14_ps(vx${ABC[N]});

      $for N in range(SIMD_TILE):
        __m512 vsqrtx${ABC[N]} = _mm512_mul_ps(vrsqrtx${ABC[N]}, vx${ABC[N]});
        __m512 vhalfrsqrtx${ABC[N]} = _mm512_mul_ps(vrsqrtx${ABC[N]}, vhalf);

      $for N in range(SIMD_TILE):
        const __m512 vresidual${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vhalfrsqrtx${ABC[N]}, vhalf);

      $for N in range(SIMD_TILE):
        vhalfrsqrtx${ABC[N]} = _mm512_fmadd_ps(vhalfrsqrtx${ABC[N]}, vresidual${ABC[N]}, vhalfrsqrtx${ABC[N]});
        vsqrtx${ABC[N]} = _mm512_fmadd_ps(vsqrtx${ABC[N]}, vresidual${ABC[N]}, vsqrtx${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m512 vadjustment${ABC[N]} = _mm512_fnmadd_ps(vsqrtx${ABC[N]}, vsqrtx${ABC[N]}, vx${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m512 vy${ABC[N]} = _mm512_fmadd_ps(vhalfrsqrtx${ABC[N]}, vadjustment${ABC[N]}, vsqrtx${ABC[N]});

      _mm512_storeu_ps(output, vy${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm512_storeu_ps(output + ${N * 16}, vy${ABC[N]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
    const __m512 vx = _mm512_loadu_ps(input);
    input += 16;

    const __m512 vrsqrtx = _mm512_rsqrt14_ps(vx);
    __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
    __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
    const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
    vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
    const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
    const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);

    _mm512_storeu_ps(output, vy);
    output += 16;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 15 * sizeof(float));
    // Prepare mask for valid 32-bit elements (depends on batch).
    batch >>= XNN_LOG2_SIZEOF_FLOAT;
    const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));

    const __m512 vx = _mm512_maskz_loadu_ps(vmask, input);
    const __m512 vrsqrtx = _mm512_maskz_rsqrt14_ps(vmask, vx);
    __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx);
    __m512 vhalfrsqrtx = _mm512_mul_ps(vrsqrtx, vhalf);
    const __m512 vresidual = _mm512_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
    vhalfrsqrtx = _mm512_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = _mm512_fmadd_ps(vsqrtx, vresidual, vsqrtx);
    const __m512 vadjustment = _mm512_fnmadd_ps(vsqrtx, vsqrtx, vx);
    const __m512 vy = _mm512_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);

    _mm512_mask_storeu_ps(output, vmask, vy);
  }
}