// 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_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);

  const __m512 vhalf = _mm512_set1_ps(params->avx512.half);
  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      const __m512 vx${ABC[0]} = _mm512_loadu_ps(x);
      $for N in range(1, SIMD_TILE):
        const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16});
      x += ${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(y, vy${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm512_storeu_ps(y + ${N * 16}, vy${ABC[N]});
      y += ${BATCH_TILE};
    }
  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
    const __m512 vx = _mm512_loadu_ps(x);
    x += 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(y, vy);
    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);
    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(y, vmask, vy);
  }
}