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

#include <immintrin.h>

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


void xnn_f32_vsqrt_ukernel__fma3_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 __m256 vhalf = _mm256_load_ps(params->fma.half);
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      const __m256 vx${ABC[0]} = _mm256_loadu_ps(x);
      $for N in range(1, SIMD_TILE):
        const __m256 vx${ABC[N]} = _mm256_loadu_ps(x + ${N * 8});
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m256 vrsqrtx${ABC[N]} = _mm256_rsqrt_ps(vx${ABC[N]});

      $for N in range(SIMD_TILE):
        __m256 vsqrtx${ABC[N]} = _mm256_mul_ps(vrsqrtx${ABC[N]}, vx${ABC[N]});
        __m256 vhalfrsqrtx${ABC[N]} = _mm256_mul_ps(vrsqrtx${ABC[N]}, vhalf);

      $for N in range(SIMD_TILE):
        const __m256 vresidual${ABC[N]} = _mm256_fnmadd_ps(vsqrtx${ABC[N]}, vhalfrsqrtx${ABC[N]}, vhalf);

      $for N in range(SIMD_TILE):
        vhalfrsqrtx${ABC[N]} = _mm256_fmadd_ps(vhalfrsqrtx${ABC[N]}, vresidual${ABC[N]}, vhalfrsqrtx${ABC[N]});
        vsqrtx${ABC[N]} = _mm256_fmadd_ps(vsqrtx${ABC[N]}, vresidual${ABC[N]}, vsqrtx${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m256 vadjustment${ABC[N]} = _mm256_fnmadd_ps(vsqrtx${ABC[N]}, vsqrtx${ABC[N]}, vx${ABC[N]});

      $for N in range(SIMD_TILE):
        const __m256 vy${ABC[N]} = _mm256_fmadd_ps(vhalfrsqrtx${ABC[N]}, vadjustment${ABC[N]}, vsqrtx${ABC[N]});

      _mm256_storeu_ps(y, vy${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm256_storeu_ps(y + ${N * 8}, vy${ABC[N]});
      y += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    const __m256 vx = _mm256_loadu_ps(x);
    x += 8;

    const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
    __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
    __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
    const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
    vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
    const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
    const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);

    _mm256_storeu_ps(y, vy);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->fma.mask_table[7] - n));

    const __m256 vx = _mm256_maskload_ps(x, vmask);

    const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
    __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
    __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
    const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
    vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
    vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
    const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
    const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);

    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (n & (4 * sizeof(float))) {
      _mm_storeu_ps(y, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      y += 4;
    }
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy_lo);
    }
  }
}